Implement LLM client and normalization settings

- Added LLMClient class for handling requests to LLM API, including methods for listing models and generating completions.
- Introduced LLMConfiguration dataclass for managing LLM configuration settings.
- Created normalization_settings module to manage normalization configurations and environment variable overrides.
- Developed JavaScript functionality for the settings interface, including model fetching and preview generation for LLM and normalization.
- Enhanced user experience with status messages and error handling in the settings UI.

.env.example

@@ -25,3 +25,14 @@ ABOGEN_TEMP_DIR=./storage/tmp
 #   id -g   # GID
 ABOGEN_UID=1000
 ABOGEN_GID=1000
+
+# Optional: Seed the web UI with working defaults for the LLM-powered
+# text normalization features. Leave these blank to configure everything
+# from the Settings page.
+ABOGEN_LLM_BASE_URL=https://localhost:11434/v1
+ABOGEN_LLM_API_KEY=ollama
+ABOGEN_LLM_MODEL=llama3.1:8b
+ABOGEN_LLM_TIMEOUT=45
+ABOGEN_LLM_CONTEXT_MODE=paragraph
+# For custom prompts, keep the text on a single line or escape newlines.
+#ABOGEN_LLM_PROMPT=You are assisting with audiobook preparation. Rewrite {{sentence}} for narration.

README.md

@@ -6,6 +6,7 @@ Abogen is a web-first text-to-speech workstation. Drop in an EPUB, PDF, Markdown
 - Natural-sounding speech powered by Kokoro-82M with per-job voice, speed, GPU toggle, and subtitle style controls
 - Clean dashboard that tracks the status, progress, and logs of every job in real time (thanks to htmx partial updates)
 - Automatic chapter detection and subtitle generation with SRT/ASS exports
+- LLM-assisted text normalization with live previews and configurable prompts
 - Runs well in Docker, ships a REST-style JSON API, and works across macOS, Linux, and Windows
 
 ## Quick start
@@ -55,6 +56,12 @@ Browse to http://localhost:8808. Uploaded source files are stored in `/data/uplo
 | `ABOGEN_TEMP_DIR` | `/data/cache` (Docker) or platform cache dir | Container path for temporary audio working files |
 | `ABOGEN_UID` | `1000` | UID that the container should run as (matches host user) |
 | `ABOGEN_GID` | `1000` | GID that the container should run as (matches host group) |
+| `ABOGEN_LLM_BASE_URL` | `""` | OpenAI-compatible endpoint used to seed the Settings → LLM panel |
+| `ABOGEN_LLM_API_KEY` | `""` | API key passed to the endpoint above |
+| `ABOGEN_LLM_MODEL` | `""` | Default model selected when you refresh the model list |
+| `ABOGEN_LLM_TIMEOUT` | `30` | Timeout (seconds) for server-side LLM requests |
+| `ABOGEN_LLM_CONTEXT_MODE` | `sentence` | Default prompt context window (`sentence`, `paragraph`, `document`) |
+| `ABOGEN_LLM_PROMPT` | `""` | Custom normalization prompt template seeded into the UI |
 
 Set any of these with `-e VAR=value` when starting the container.
 
@@ -123,6 +130,15 @@ Abogen falls back to CPU rendering if no GPU is available.
 
 Multiple jobs can run sequentially; the worker processes them in order.
 
+## LLM-assisted text normalization
+Abogen can hand tricky apostrophes and contractions to an OpenAI-compatible large language model. Configure it from **Settings → LLM**:
+
+1. Enter the base URL for your endpoint (Ollama, OpenAI proxy, etc.) and an API key if required.
+2. Click **Refresh models** to load the catalog, pick a default model, and adjust the timeout or prompt template.
+3. Use the preview box to test the prompt, then save the settings. The Normalization panel can synthesize a short audio preview with the current configuration.
+
+When you are running inside Docker or a CI pipeline, seed the form automatically with `ABOGEN_LLM_*` variables in your `.env` file. The `.env.example` file includes sample values for a local Ollama server.
+
 ## JSON endpoints
 Need machine-readable status updates? The dashboard calls a small set of helper endpoints you can reuse:
 - `GET /api/jobs/<id>` returns job metadata, progress, and log lines in JSON.
@@ -138,6 +154,7 @@ Most behaviour is controlled through the UI, but a few environment variables are
 - `ABOGEN_SETTINGS_DIR` – change where Abogen stores its JSON settings/configuration files.
 - `ABOGEN_TEMP_DIR` – change where temporary uploads and cache files are stored.
 - `ABOGEN_OUTPUT_DIR` – change where rendered audio/subtitles are written.
+- `ABOGEN_LLM_*` – seed the Settings → LLM panel with defaults for base URL, API key, model, timeout, prompt, and context mode.
 
 If unset, Abogen picks sensible defaults suitable for local usage.
 

abogen/chunking.py

@@ -7,6 +7,7 @@ from typing import Pattern
 import re
 
 from abogen.kokoro_text_normalization import ApostropheConfig, normalize_for_pipeline
+from abogen.normalization_settings import build_apostrophe_config, get_runtime_settings
 
 ChunkLevel = Literal["paragraph", "sentence"]
 
@@ -78,7 +79,9 @@ def _normalize_whitespace(value: str) -> str:
 
 
 def _normalize_chunk_text(value: str) -> str:
-    normalized = normalize_for_pipeline(value, config=_PIPELINE_APOSTROPHE_CONFIG)
+    settings = get_runtime_settings()
+    config = build_apostrophe_config(settings=settings, base=_PIPELINE_APOSTROPHE_CONFIG)
+    normalized = normalize_for_pipeline(value, config=config, settings=settings)
     return _normalize_whitespace(normalized)
 
 

abogen/kokoro_text_normalization.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 import re
 import unicodedata
 from dataclasses import dataclass
-from typing import Callable, Iterable, List, Optional, Sequence, Tuple
+from typing import Any, Callable, Dict, Iterable, List, Mapping, Optional, Sequence, Tuple
 try:  # pragma: no cover - optional dependency guard
     from num2words import num2words
 except Exception:  # pragma: no cover - graceful degradation
@@ -598,21 +598,152 @@ def apply_phoneme_hints(text: str, iz_marker="‹IZ›") -> str:
 DEFAULT_APOSTROPHE_CONFIG = ApostropheConfig()
 
 
-def normalize_for_pipeline(text: str, *, config: Optional[ApostropheConfig] = None) -> str:
+_MUSTACHE_PATTERN = re.compile(r"{{\s*([a-zA-Z0-9_]+)\s*}}")
-    """Normalize text for the synthesis pipeline.
+_LLM_SYSTEM_PROMPT = (
+    "You rewrite text for audiobook narration. Expand or clarify contractions and apostrophes "
+    "so the output is explicit and natural to read aloud. Respond with only the rewritten text."
+)
 
-    This expands contractions, normalizes apostrophes, and adds phoneme hints
-    using the provided configuration so downstream chunking and synthesis share
-    the same representation.
-    """
 
-    cfg = config or DEFAULT_APOSTROPHE_CONFIG
+def _render_mustache(template: str, context: Mapping[str, str]) -> str:
-    normalized, _details = normalize_apostrophes(text, cfg)
+    if not template:
-    normalized = expand_titles_and_suffixes(normalized)
+        return ""
-    normalized = ensure_terminal_punctuation(normalized)
+
+    def _replace(match: re.Match[str]) -> str:
+        key = match.group(1)
+        return context.get(key, "")
+
+    return _MUSTACHE_PATTERN.sub(_replace, template)
+
+
+_SENTENCE_CAPTURE_RE = re.compile(r"[^.!?]+[.!?]+|[^.!?]+$", re.MULTILINE)
+
+
+def _split_sentences_for_llm(text: str) -> List[str]:
+    sentences = [segment.strip() for segment in _SENTENCE_CAPTURE_RE.findall(text or "")]
+    return [segment for segment in sentences if segment]
+
+
+def _normalize_with_llm(
+    text: str,
+    *,
+    settings: Mapping[str, Any],
+    config: ApostropheConfig,
+) -> str:
+    from abogen.normalization_settings import build_llm_configuration, DEFAULT_LLM_PROMPT
+    from abogen.llm_client import generate_completion, LLMClientError
+
+    llm_config = build_llm_configuration(settings)
+    if not llm_config.is_configured():
+        raise LLMClientError("LLM configuration is incomplete")
+
+    prompt_template = str(settings.get("llm_prompt") or DEFAULT_LLM_PROMPT)
+    context_mode = str(settings.get("llm_context_mode") or "sentence").strip().lower()
+    lines = text.splitlines(keepends=True)
+    if not lines:
+        return text
+
+    normalized_lines: List[str] = []
+    for raw_line in lines:
+        newline = ""
+        if raw_line.endswith(("\r", "\n")):
+            stripped_newline = raw_line.rstrip("\r\n")
+            newline = raw_line[len(stripped_newline):]
+            line_body = stripped_newline
+        else:
+            line_body = raw_line
+
+        if not line_body.strip():
+            normalized_lines.append(line_body + newline)
+            continue
+
+        leading_ws = line_body[: len(line_body) - len(line_body.lstrip())]
+        trailing_ws = line_body[len(line_body.rstrip()):]
+        core = line_body[len(leading_ws) : len(line_body) - len(trailing_ws)]
+
+        paragraph_context = core
+        if context_mode == "sentence":
+            sentences = _split_sentences_for_llm(core)
+            if not sentences:
+                normalized_lines.append(line_body + newline)
+                continue
+            rewritten_sentences: List[str] = []
+            for sentence in sentences:
+                prompt_context = {
+                    "text": sentence,
+                    "sentence": sentence,
+                    "paragraph": paragraph_context,
+                }
+                prompt = _render_mustache(prompt_template, prompt_context)
+                completion = generate_completion(
+                    llm_config,
+                    system_message=_LLM_SYSTEM_PROMPT,
+                    user_message=prompt,
+                )
+                rewritten_sentences.append(completion.strip())
+            normalized_core = " ".join(filter(None, rewritten_sentences)) or core
+        else:
+            prompt_context = {
+                "text": core,
+                "sentence": core,
+                "paragraph": paragraph_context,
+            }
+            prompt = _render_mustache(prompt_template, prompt_context)
+            normalized_core = generate_completion(
+                llm_config,
+                system_message=_LLM_SYSTEM_PROMPT,
+                user_message=prompt,
+            ).strip() or core
+
+        rebuilt = f"{leading_ws}{normalized_core}{trailing_ws}{newline}"
+        normalized_lines.append(rebuilt)
+
+    result = "".join(normalized_lines)
+    return result if result else text
+
+
+def normalize_for_pipeline(
+    text: str,
+    *,
+    config: Optional[ApostropheConfig] = None,
+    settings: Optional[Mapping[str, Any]] = None,
+) -> str:
+    """Normalize text for the synthesis pipeline with runtime settings."""
+
+    from abogen.normalization_settings import build_apostrophe_config, get_runtime_settings
+    from abogen.llm_client import LLMClientError
+
+    runtime_settings = settings or get_runtime_settings()
+    base_config = config or DEFAULT_APOSTROPHE_CONFIG
+    cfg = build_apostrophe_config(settings=runtime_settings, base=base_config)
+
+    mode = str(runtime_settings.get("normalization_apostrophe_mode", "spacy")).lower()
+    normalized = text
+
+    if mode == "off":
+        normalized = normalize_unicode_apostrophes(text)
+        if cfg.convert_numbers:
             normalized = _normalize_grouped_numbers(normalized, cfg)
+        normalized = _cleanup_spacing(normalized)
+    elif mode == "llm":
+        try:
+            normalized = _normalize_with_llm(text, settings=runtime_settings, config=cfg)
+        except LLMClientError:
+            raise
+        if cfg.convert_numbers:
+            normalized = _normalize_grouped_numbers(normalized, cfg)
+        normalized = _cleanup_spacing(normalized)
+    else:
+        normalized, _ = normalize_apostrophes(text, cfg)
+
+    if runtime_settings.get("normalization_titles", True):
+        normalized = expand_titles_and_suffixes(normalized)
+    if runtime_settings.get("normalization_terminal", True):
+        normalized = ensure_terminal_punctuation(normalized)
+
     if cfg.add_phoneme_hints:
         normalized = apply_phoneme_hints(normalized, iz_marker=cfg.sibilant_iz_marker)
+
     return normalized
 
 # ---------- Example Usage ----------

abogen/llm_client.py

@@ -0,0 +1,148 @@
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from typing import Any, Dict, List, Mapping, Optional
+from urllib import error, parse, request
+
+
+class LLMClientError(RuntimeError):
+    """Raised when an LLM request fails."""
+
+
+@dataclass(frozen=True)
+class LLMConfiguration:
+    base_url: str
+    api_key: str
+    model: str
+    timeout: float = 30.0
+
+    def is_configured(self) -> bool:
+        return bool(self.base_url.strip() and self.model.strip())
+
+
+_DEFAULT_HEADERS = {
+    "Content-Type": "application/json",
+    "Accept": "application/json",
+}
+
+
+def _normalized_base_url(base_url: str) -> str:
+    trimmed = (base_url or "").strip()
+    if not trimmed:
+        raise LLMClientError("LLM base URL is required")
+    if not trimmed.endswith("/"):
+        trimmed += "/"
+    return trimmed
+
+
+def _build_url(base_url: str, path: str) -> str:
+    normalized = _normalized_base_url(base_url)
+    return parse.urljoin(normalized, path.lstrip("/"))
+
+
+def _build_headers(api_key: str) -> Dict[str, str]:
+    headers = dict(_DEFAULT_HEADERS)
+    token = (api_key or "").strip()
+    if token and token.lower() != "ollama":
+        headers["Authorization"] = f"Bearer {token}"
+    return headers
+
+
+def _perform_request(
+    method: str,
+    url: str,
+    *,
+    headers: Optional[Mapping[str, str]] = None,
+    payload: Optional[Mapping[str, Any]] = None,
+    timeout: float = 30.0,
+) -> Any:
+    data_bytes: Optional[bytes] = None
+    if payload is not None:
+        data_bytes = json.dumps(payload).encode("utf-8")
+    request_headers = dict(headers or {})
+    req = request.Request(url, data=data_bytes, headers=request_headers, method=method.upper())
+    try:
+        with request.urlopen(req, timeout=timeout) as response:
+            body = response.read()
+    except error.HTTPError as exc:  # pragma: no cover - defensive network guard
+        message = exc.read().decode("utf-8", "ignore") if exc.fp else exc.reason
+        raise LLMClientError(f"LLM request failed ({exc.code}): {message}") from exc
+    except error.URLError as exc:  # pragma: no cover - defensive network guard
+        raise LLMClientError(f"LLM request failed: {exc.reason}") from exc
+    except Exception as exc:  # pragma: no cover - defensive network guard
+        raise LLMClientError("LLM request failed") from exc
+
+    if not body:
+        return None
+    try:
+        return json.loads(body.decode("utf-8"))
+    except json.JSONDecodeError as exc:
+        raise LLMClientError("LLM response was not valid JSON") from exc
+
+
+def list_models(configuration: LLMConfiguration) -> List[Dict[str, str]]:
+    if not configuration.is_configured() and not configuration.base_url.strip():
+        raise LLMClientError("LLM configuration is incomplete")
+    url = _build_url(configuration.base_url, "v1/models")
+    headers = _build_headers(configuration.api_key)
+    payload = _perform_request("GET", url, headers=headers, timeout=configuration.timeout)
+    if not isinstance(payload, Mapping):
+        raise LLMClientError("Unexpected response when listing models")
+    data = payload.get("data")
+    if not isinstance(data, list):
+        return []
+    models: List[Dict[str, str]] = []
+    for entry in data:
+        if not isinstance(entry, Mapping):
+            continue
+        identifier = str(entry.get("id") or "").strip()
+        if not identifier:
+            continue
+        description = str(entry.get("name") or entry.get("description") or identifier)
+        models.append({"id": identifier, "label": description})
+    return models
+
+
+def generate_completion(
+    configuration: LLMConfiguration,
+    *,
+    system_message: str,
+    user_message: str,
+    temperature: float = 0.2,
+    max_tokens: Optional[int] = None,
+) -> str:
+    if not configuration.is_configured():
+        raise LLMClientError("LLM configuration is incomplete")
+
+    url = _build_url(configuration.base_url, "v1/chat/completions")
+    headers = _build_headers(configuration.api_key)
+    payload: Dict[str, Any] = {
+        "model": configuration.model,
+        "messages": [
+            {"role": "system", "content": system_message},
+            {"role": "user", "content": user_message},
+        ],
+        "temperature": temperature,
+    }
+    if max_tokens is not None:
+        payload["max_tokens"] = max_tokens
+
+    response = _perform_request("POST", url, headers=headers, payload=payload, timeout=configuration.timeout)
+    if not isinstance(response, Mapping):
+        raise LLMClientError("Unexpected response from LLM")
+    choices = response.get("choices")
+    if not isinstance(choices, list) or not choices:
+        raise LLMClientError("LLM response did not include choices")
+    first = choices[0]
+    if not isinstance(first, Mapping):
+        raise LLMClientError("LLM response choice was invalid")
+    message = first.get("message")
+    if isinstance(message, Mapping):
+        content = message.get("content")
+        if isinstance(content, str) and content.strip():
+            return content.strip()
+    text = first.get("text")
+    if isinstance(text, str) and text.strip():
+        return text.strip()
+    raise LLMClientError("LLM response did not include text content")

abogen/normalization_settings.py

@@ -0,0 +1,151 @@
+from __future__ import annotations
+
+import os
+from dataclasses import replace
+from functools import lru_cache
+from typing import Any, Dict, Mapping, Optional
+
+from abogen.kokoro_text_normalization import ApostropheConfig
+from abogen.llm_client import LLMConfiguration
+from abogen.utils import load_config
+
+DEFAULT_LLM_PROMPT = (
+    "You are assisting with audiobook preparation. Rewrite the provided sentence so apostrophes and "
+    "contractions are unambiguous for text-to-speech. Respond with only the rewritten sentence.\n"
+    "Sentence: {{ sentence }}\n"
+    "Context: {{ paragraph }}"
+)
+
+_SETTINGS_DEFAULTS: Dict[str, Any] = {
+    "llm_base_url": "",
+    "llm_api_key": "",
+    "llm_model": "",
+    "llm_timeout": 30.0,
+    "llm_prompt": DEFAULT_LLM_PROMPT,
+    "llm_context_mode": "sentence",
+    "normalization_numbers": True,
+    "normalization_titles": True,
+    "normalization_terminal": True,
+    "normalization_phoneme_hints": True,
+    "normalization_apostrophe_mode": "spacy",
+}
+
+_ENVIRONMENT_KEYS: Dict[str, str] = {
+    "llm_base_url": "ABOGEN_LLM_BASE_URL",
+    "llm_api_key": "ABOGEN_LLM_API_KEY",
+    "llm_model": "ABOGEN_LLM_MODEL",
+    "llm_timeout": "ABOGEN_LLM_TIMEOUT",
+    "llm_prompt": "ABOGEN_LLM_PROMPT",
+    "llm_context_mode": "ABOGEN_LLM_CONTEXT_MODE",
+}
+
+NORMALIZATION_SAMPLE_TEXTS: Dict[str, str] = {
+    "apostrophes": "I've heard the captain'll arrive by dusk, but they'd said the same yesterday.",
+    "numbers": "The ledger listed 1,204 outstanding debts totaling $57,890.",
+    "titles": "Dr. Smith met Mr. O'Leary outside St. John's Church on Jan. 4th.",
+    "punctuation": "Meet me at the docks tonight We'll decide then",  # missing punctuation
+}
+
+
+@lru_cache(maxsize=1)
+def _environment_defaults() -> Dict[str, Any]:
+    overrides: Dict[str, Any] = {}
+    for key, env_var in _ENVIRONMENT_KEYS.items():
+        default = _SETTINGS_DEFAULTS.get(key)
+        if default is None:
+            continue
+        value = os.environ.get(env_var)
+        if value is None or value == "":
+            continue
+        if isinstance(default, bool):
+            overrides[key] = _coerce_bool(value, default)
+        elif isinstance(default, float):
+            overrides[key] = _coerce_float(value, float(default))
+        else:
+            overrides[key] = value
+    return overrides
+
+
+def environment_llm_defaults() -> Dict[str, Any]:
+    return dict(_environment_defaults())
+
+
+def _coerce_bool(value: Any, default: bool) -> bool:
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        lowered = value.strip().lower()
+        if lowered in {"1", "true", "yes", "on"}:
+            return True
+        if lowered in {"0", "false", "no", "off"}:
+            return False
+    return default
+
+
+def _coerce_float(value: Any, default: float) -> float:
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        return default
+
+
+def _extract_settings(source: Mapping[str, Any]) -> Dict[str, Any]:
+    env_defaults = _environment_defaults()
+    extracted: Dict[str, Any] = {}
+    for key, default in _SETTINGS_DEFAULTS.items():
+        if key in source:
+            raw_value = source.get(key)
+        elif key in env_defaults:
+            raw_value = env_defaults[key]
+        else:
+            raw_value = default
+        if isinstance(default, bool):
+            extracted[key] = _coerce_bool(raw_value, default)
+        elif isinstance(default, float):
+            extracted[key] = _coerce_float(raw_value, default)
+        else:
+            extracted[key] = str(raw_value or "") if isinstance(default, str) else raw_value
+    return extracted
+
+
+@lru_cache(maxsize=1)
+def _cached_settings() -> Dict[str, Any]:
+    config = load_config() or {}
+    return _extract_settings(config)
+
+
+def get_runtime_settings() -> Dict[str, Any]:
+    return dict(_cached_settings())
+
+
+def clear_cached_settings() -> None:
+    _cached_settings.cache_clear()
+
+
+def build_apostrophe_config(
+    *,
+    settings: Mapping[str, Any],
+    base: Optional[ApostropheConfig] = None,
+) -> ApostropheConfig:
+    config = replace(base or ApostropheConfig())
+    config.convert_numbers = bool(settings.get("normalization_numbers", True))
+    config.add_phoneme_hints = bool(settings.get("normalization_phoneme_hints", True))
+    return config
+
+
+def build_llm_configuration(settings: Mapping[str, Any]) -> LLMConfiguration:
+    return LLMConfiguration(
+        base_url=str(settings.get("llm_base_url") or ""),
+        api_key=str(settings.get("llm_api_key") or ""),
+        model=str(settings.get("llm_model") or ""),
+        timeout=_coerce_float(settings.get("llm_timeout"), float(_SETTINGS_DEFAULTS["llm_timeout"])),
+    )
+
+
+def apply_overrides(base: Mapping[str, Any], overrides: Mapping[str, Any]) -> Dict[str, Any]:
+    merged: Dict[str, Any] = dict(base)
+    for key, value in overrides.items():
+        if key not in _SETTINGS_DEFAULTS:
+            continue
+        merged[key] = value
+    return merged

abogen/web/conversion_runner.py

@@ -21,6 +21,11 @@ import static_ffmpeg
 from abogen.constants import VOICES_INTERNAL
 from abogen.epub3.exporter import build_epub3_package
 from abogen.kokoro_text_normalization import ApostropheConfig, normalize_for_pipeline
+from abogen.normalization_settings import (
+    build_apostrophe_config,
+    build_llm_configuration,
+    get_runtime_settings,
+)
 from abogen.entity_analysis import normalize_token as normalize_entity_token
 from abogen.text_extractor import ExtractedChapter, extract_from_path
 from abogen.utils import (
@@ -34,6 +39,8 @@ from abogen.utils import (
 )
 from abogen.voice_cache import ensure_voice_assets
 from abogen.voice_formulas import extract_voice_ids, get_new_voice
+from abogen.pronunciation_store import increment_usage
+from abogen.llm_client import LLMClientError
 
 from .service import Job, JobStatus
 
@@ -460,17 +467,13 @@ def _merge_metadata(
 _APOSTROPHE_CONFIG = ApostropheConfig()
 
 
-def _normalize_for_pipeline(text: str) -> str:
-    return normalize_for_pipeline(text, config=_APOSTROPHE_CONFIG)
-
-
 def _compile_pronunciation_rules(
     overrides: Optional[Iterable[Mapping[str, Any]]],
-) -> List[tuple[re.Pattern[str], str]]:
+) -> List[Dict[str, Any]]:
     if not overrides:
         return []
 
-    candidates: List[tuple[str, str]] = []
+    candidates: List[Dict[str, Any]] = []
     seen: set[str] = set()
 
     for entry in overrides:
@@ -499,34 +502,64 @@ def _compile_pronunciation_rules(
         if not token_values:
             continue
 
+        usage_normalized = str(entry.get("normalized") or "").strip()
+        if not usage_normalized and token_values:
+            usage_normalized = normalize_entity_token(token_values[0]) or token_values[0]
+        usage_token = str(entry.get("token") or token_values[0])
+
         for token_value in token_values:
             key = token_value.casefold()
             if key in seen:
                 continue
             seen.add(key)
-            candidates.append((token_value, pronunciation_value))
+            candidates.append(
+                {
+                    "token": token_value,
+                    "normalized": usage_normalized,
+                    "replacement": pronunciation_value,
+                }
+            )
 
     if not candidates:
         return []
 
-    candidates.sort(key=lambda item: len(item[0]), reverse=True)
+    candidates.sort(key=lambda item: len(item["token"]), reverse=True)
-    compiled: List[tuple[re.Pattern[str], str]] = []
+    compiled: List[Dict[str, Any]] = []
-    for token_value, pronunciation_value in candidates:
+    for candidate in candidates:
+        token_value = candidate["token"]
+        pronunciation_value = candidate["replacement"]
         escaped = re.escape(token_value)
         pattern = re.compile(rf"(?i)(?<!\w){escaped}(?P<possessive>'s|\u2019s|\u2019)?(?!\w)")
-        compiled.append((pattern, pronunciation_value))
+        compiled.append(
+            {
+                "pattern": pattern,
+                "replacement": pronunciation_value,
+                "normalized": candidate.get("normalized") or token_value,
+                "token": candidate.get("token") or token_value,
+            }
+        )
 
     return compiled
 
 
-def _apply_pronunciation_rules(text: str, rules: List[tuple[re.Pattern[str], str]]) -> str:
+def _apply_pronunciation_rules(
+    text: str,
+    rules: List[Dict[str, Any]],
+    usage_counter: Optional[Dict[str, int]] = None,
+) -> str:
     if not text or not rules:
         return text
 
     result = text
-    for pattern, pronunciation_value in rules:
+    for rule in rules:
+        pattern = rule["pattern"]
+        pronunciation_value = rule["replacement"]
+        usage_key = str(rule.get("normalized") or "").strip()
+
         def _replacement(match: re.Match[str]) -> str:
             suffix = match.group("possessive") or ""
+            if usage_counter is not None and usage_key:
+                usage_counter[usage_key] = usage_counter.get(usage_key, 0) + 1
             return pronunciation_value + suffix
 
         result = pattern.sub(_replacement, result)
@@ -604,6 +637,25 @@ def _group_chunks_by_chapter(chunks: Iterable[Dict[str, Any]]) -> Dict[int, List
     return grouped
 
 
+def _record_override_usage(
+    job: Job,
+    usage_counter: Mapping[str, int],
+    token_map: Mapping[str, str],
+) -> None:
+    if not usage_counter:
+        return
+
+    language = getattr(job, "language", "") or "a"
+    for normalized, amount in usage_counter.items():
+        if amount <= 0:
+            continue
+        token_value = token_map.get(normalized, normalized)
+        try:
+            increment_usage(language=language, token=token_value, amount=int(amount))
+        except Exception:  # pragma: no cover - defensive logging
+            job.add_log(f"Failed to record usage for override {token_value}", level="warning")
+
+
 def _safe_int(value: Any, default: int = 0) -> int:
     try:
         return int(value)
@@ -846,6 +898,19 @@ def run_conversion_job(job: Job) -> None:
     job.add_log("Preparing conversion pipeline")
     canceller = _make_canceller(job)
 
+    normalization_settings = get_runtime_settings()
+    apostrophe_config = build_apostrophe_config(
+        settings=normalization_settings,
+        base=_APOSTROPHE_CONFIG,
+    )
+    apostrophe_mode = str(normalization_settings.get("normalization_apostrophe_mode", "spacy")).lower()
+    if apostrophe_mode == "llm":
+        llm_config = build_llm_configuration(normalization_settings)
+        if not llm_config.is_configured():
+            raise RuntimeError(
+                "LLM-based apostrophe normalization is selected, but the LLM configuration is incomplete."
+            )
+
     sink_stack = ExitStack()
     subtitle_writer: Optional[SubtitleWriter] = None
     chapter_paths: list[Path] = []
@@ -857,6 +922,8 @@ def run_conversion_job(job: Job) -> None:
     pipeline: Any = None
     chunk_groups: Dict[int, List[Dict[str, Any]]] = {}
     active_chapter_configs: List[Dict[str, Any]] = []
+    usage_counter: Dict[str, int] = defaultdict(int)
+    override_token_map: Dict[str, str] = {}
     try:
         pipeline = _load_pipeline(job)
         _initialize_voice_cache(job)
@@ -869,6 +936,15 @@ def run_conversion_job(job: Job) -> None:
                 f"Applying {count} pronunciation override{'s' if count != 1 else ''} during conversion.",
                 level="debug",
             )
+        for override_entry in job.pronunciation_overrides or []:
+            if not isinstance(override_entry, Mapping):
+                continue
+            raw_token = str(override_entry.get("token") or "").strip()
+            normalized_value = str(override_entry.get("normalized") or "").strip()
+            if not normalized_value and raw_token:
+                normalized_value = normalize_entity_token(raw_token) or raw_token
+            if normalized_value:
+                override_token_map.setdefault(normalized_value, raw_token or normalized_value)
 
         if not job.chapters:
             filtered, skipped_info = _auto_select_relevant_chapters(extraction.chapters, file_type)
@@ -986,8 +1062,20 @@ def run_conversion_job(job: Job) -> None:
             nonlocal processed_chars, subtitle_index, current_time
             source_text = str(text or "")
             if pronunciation_rules:
-                source_text = _apply_pronunciation_rules(source_text, pronunciation_rules)
+                source_text = _apply_pronunciation_rules(
-            normalized = _normalize_for_pipeline(source_text)
+                    source_text,
+                    pronunciation_rules,
+                    usage_counter,
+                )
+            try:
+                normalized = normalize_for_pipeline(
+                    source_text,
+                    config=apostrophe_config,
+                    settings=normalization_settings,
+                )
+            except LLMClientError as exc:
+                job.add_log(f"LLM normalization failed: {exc}", level="error")
+                raise
             local_segments = 0
 
             for segment in pipeline(
@@ -1278,6 +1366,9 @@ def run_conversion_job(job: Job) -> None:
             "generate_epub3": job.generate_epub3,
         }
 
+        if usage_counter:
+            _record_override_usage(job, usage_counter, override_token_map)
+
         if metadata_dir:
             metadata_dir.mkdir(parents=True, exist_ok=True)
             metadata_file = metadata_dir / "metadata.json"

abogen/web/routes.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import base64
 import io
 import json
 import math
@@ -46,6 +47,17 @@ from abogen.constants import (
     VOICES_INTERNAL,
 )
 from abogen.kokoro_text_normalization import normalize_for_pipeline, normalize_roman_numeral_titles
+from abogen.normalization_settings import (
+    DEFAULT_LLM_PROMPT,
+    NORMALIZATION_SAMPLE_TEXTS,
+    apply_overrides as apply_normalization_overrides,
+    build_apostrophe_config,
+    build_llm_configuration,
+    clear_cached_settings,
+    environment_llm_defaults,
+    get_runtime_settings,
+)
+from abogen.llm_client import LLMClientError, LLMConfiguration, generate_completion, list_models
 from abogen.utils import (
     calculate_text_length,
     get_user_output_path,
@@ -1853,6 +1865,17 @@ SAVE_MODE_LABELS = {
 
 LEGACY_SAVE_MODE_MAP = {label: key for key, label in SAVE_MODE_LABELS.items()}
 
+_APOSTROPHE_MODE_OPTIONS = [
+    {"value": "off", "label": "Off"},
+    {"value": "spacy", "label": "spaCy (built-in)"},
+    {"value": "llm", "label": "LLM assisted"},
+]
+
+_LLM_CONTEXT_OPTIONS = [
+    {"value": "sentence", "label": "Sentence only"},
+    {"value": "paragraph", "label": "Sentence with paragraph context"},
+]
+
 BOOLEAN_SETTINGS = {
     "replace_single_newlines",
     "use_gpu",
@@ -1862,9 +1885,13 @@ BOOLEAN_SETTINGS = {
     "generate_epub3",
     "enable_entity_recognition",
     "auto_prefix_chapter_titles",
+    "normalization_numbers",
+    "normalization_titles",
+    "normalization_terminal",
+    "normalization_phoneme_hints",
 }
 
-FLOAT_SETTINGS = {"silence_between_chapters", "chapter_intro_delay"}
+FLOAT_SETTINGS = {"silence_between_chapters", "chapter_intro_delay", "llm_timeout"}
 INT_SETTINGS = {"max_subtitle_words", "speaker_analysis_threshold"}
 
 
@@ -1873,6 +1900,7 @@ def _has_output_override() -> bool:
 
 
 def _settings_defaults() -> Dict[str, Any]:
+    llm_env_defaults = environment_llm_defaults()
     return {
         "output_format": "wav",
         "subtitle_format": "srt",
@@ -1894,9 +1922,39 @@ def _settings_defaults() -> Dict[str, Any]:
         "speaker_analysis_threshold": _DEFAULT_ANALYSIS_THRESHOLD,
         "speaker_pronunciation_sentence": "This is {{name}} speaking.",
         "speaker_random_languages": [],
+    "llm_base_url": llm_env_defaults.get("llm_base_url", ""),
+    "llm_api_key": llm_env_defaults.get("llm_api_key", ""),
+    "llm_model": llm_env_defaults.get("llm_model", ""),
+    "llm_timeout": llm_env_defaults.get("llm_timeout", 30.0),
+    "llm_prompt": llm_env_defaults.get("llm_prompt", DEFAULT_LLM_PROMPT),
+    "llm_context_mode": llm_env_defaults.get("llm_context_mode", "sentence"),
+        "normalization_numbers": True,
+        "normalization_titles": True,
+        "normalization_terminal": True,
+        "normalization_phoneme_hints": True,
+        "normalization_apostrophe_mode": "spacy",
     }
 
 
+def _llm_ready(settings: Mapping[str, Any]) -> bool:
+    base_url = str(settings.get("llm_base_url") or "").strip()
+    return bool(base_url)
+
+
+_PROMPT_TOKEN_RE = re.compile(r"{{\s*([a-zA-Z0-9_]+)\s*}}")
+
+
+def _render_prompt_template(template: str, context: Mapping[str, str]) -> str:
+    if not template:
+        return ""
+
+    def _replace(match: re.Match[str]) -> str:
+        key = match.group(1)
+        return context.get(key, "")
+
+    return _PROMPT_TOKEN_RE.sub(_replace, template)
+
+
 def _coerce_bool(value: Any, default: bool) -> bool:
     if isinstance(value, bool):
         return value
@@ -1959,6 +2017,23 @@ def _normalize_setting_value(key: str, value: Any, defaults: Dict[str, Any]) ->
         if isinstance(value, str) and value in _CHUNK_LEVEL_VALUES:
             return value
         return defaults[key]
+    if key == "normalization_apostrophe_mode":
+        if isinstance(value, str):
+            normalized_mode = value.strip().lower()
+            if normalized_mode in {"off", "spacy", "llm"}:
+                return normalized_mode
+        return defaults[key]
+    if key == "llm_context_mode":
+        if isinstance(value, str):
+            normalized_scope = value.strip().lower()
+            if normalized_scope in {"sentence", "paragraph"}:
+                return normalized_scope
+        return defaults[key]
+    if key == "llm_prompt":
+        candidate = str(value or "").strip()
+        return candidate if candidate else defaults[key]
+    if key in {"llm_base_url", "llm_api_key", "llm_model"}:
+        return str(value or "").strip()
     if key == "speaker_random_languages":
         if isinstance(value, (list, tuple, set)):
             return [code for code in value if isinstance(code, str) and code in LANGUAGE_DESCRIPTIONS]
@@ -2100,6 +2175,68 @@ def _get_preview_pipeline(language: str, device: str):
         return pipeline
 
 
+def _synthesize_audio_from_normalized(
+    *,
+    normalized_text: str,
+    voice_spec: str,
+    language: str,
+    speed: float,
+    use_gpu: bool,
+    max_seconds: float,
+) -> np.ndarray:
+    if not normalized_text.strip():
+        raise ValueError("Preview text is required")
+
+    device = "cpu"
+    if use_gpu:
+        try:
+            device = _select_device()
+        except Exception:
+            device = "cpu"
+            use_gpu = False
+
+    pipeline = _get_preview_pipeline(language, device)
+    if pipeline is None:
+        raise RuntimeError("Preview pipeline is unavailable")
+
+    voice_choice: Any = voice_spec
+    if voice_spec and "*" in voice_spec:
+        voice_choice = get_new_voice(pipeline, voice_spec, use_gpu)
+
+    segments = pipeline(
+        normalized_text,
+        voice=voice_choice,
+        speed=speed,
+        split_pattern=SPLIT_PATTERN,
+    )
+
+    audio_chunks: List[np.ndarray] = []
+    accumulated = 0
+    max_samples = int(max(1.0, max_seconds) * SAMPLE_RATE)
+
+    for segment in segments:
+        graphemes = getattr(segment, "graphemes", "").strip()
+        if not graphemes:
+            continue
+        audio = _to_float32(getattr(segment, "audio", None))
+        if audio.size == 0:
+            continue
+        remaining = max_samples - accumulated
+        if remaining <= 0:
+            break
+        if audio.shape[0] > remaining:
+            audio = audio[:remaining]
+        audio_chunks.append(audio)
+        accumulated += audio.shape[0]
+        if accumulated >= max_samples:
+            break
+
+    if not audio_chunks:
+        raise RuntimeError("Preview could not be generated")
+
+    return np.concatenate(audio_chunks)
+
+
 @web_bp.app_template_filter("datetimeformat")
 def datetimeformat(value: float, fmt: str = "%Y-%m-%d %H:%M:%S") -> str:
     if not value:
@@ -2192,9 +2329,28 @@ def settings_page() -> ResponseReturnValue:
         ]
         updated["speaker_random_languages"] = random_languages
 
+        updated["llm_base_url"] = _normalize_setting_value(
+            "llm_base_url", form.get("llm_base_url"), defaults
+        )
+        updated["llm_api_key"] = _normalize_setting_value(
+            "llm_api_key", form.get("llm_api_key"), defaults
+        )
+        updated["llm_model"] = _normalize_setting_value("llm_model", form.get("llm_model"), defaults)
+        updated["llm_prompt"] = _normalize_setting_value("llm_prompt", form.get("llm_prompt"), defaults)
+        updated["llm_context_mode"] = _normalize_setting_value(
+            "llm_context_mode", form.get("llm_context_mode"), defaults
+        )
+        updated["llm_timeout"] = _normalize_setting_value("llm_timeout", form.get("llm_timeout"), defaults)
+        updated["normalization_apostrophe_mode"] = _normalize_setting_value(
+            "normalization_apostrophe_mode",
+            form.get("normalization_apostrophe_mode"),
+            defaults,
+        )
+
         cfg = load_config() or {}
         cfg.update(updated)
         save_config(cfg)
+        clear_cached_settings()
         return redirect(url_for("web.settings_page", saved="1"))
 
     save_locations = [
@@ -2206,10 +2362,174 @@ def settings_page() -> ResponseReturnValue:
         "save_locations": save_locations,
         "default_output_dir": get_user_output_path(),
         "saved": request.args.get("saved") == "1",
+        "apostrophe_modes": _APOSTROPHE_MODE_OPTIONS,
+        "llm_context_options": _LLM_CONTEXT_OPTIONS,
+        "llm_ready": _llm_ready(current_settings),
+        "normalization_samples": NORMALIZATION_SAMPLE_TEXTS,
     }
     return render_template("settings.html", **context)
 
 
+@api_bp.post("/llm/models")
+def api_llm_models() -> ResponseReturnValue:
+    payload = request.get_json(force=True, silent=False) or {}
+    current_settings = get_runtime_settings()
+
+    base_url = str(payload.get("base_url") or payload.get("llm_base_url") or current_settings.get("llm_base_url") or "").strip()
+    if not base_url:
+        return jsonify({"error": "LLM base URL is required."}), 400
+
+    api_key = str(payload.get("api_key") or payload.get("llm_api_key") or current_settings.get("llm_api_key") or "")
+    timeout = _coerce_float(payload.get("timeout"), current_settings.get("llm_timeout", 30.0))
+
+    overrides = {
+        "llm_base_url": base_url,
+        "llm_api_key": api_key,
+        "llm_timeout": timeout,
+    }
+
+    merged = apply_normalization_overrides(current_settings, overrides)
+    configuration = build_llm_configuration(merged)
+    try:
+        models = list_models(configuration)
+    except LLMClientError as exc:
+        return jsonify({"error": str(exc)}), 400
+    return jsonify({"models": models})
+
+
+@api_bp.post("/llm/preview")
+def api_llm_preview() -> ResponseReturnValue:
+    payload = request.get_json(force=True, silent=False) or {}
+    sample_text = str(payload.get("text") or "").strip()
+    if not sample_text:
+        return jsonify({"error": "Text is required."}), 400
+
+    base_settings = get_runtime_settings()
+    overrides: Dict[str, Any] = {
+        "llm_base_url": str(
+            payload.get("base_url")
+            or payload.get("llm_base_url")
+            or base_settings.get("llm_base_url")
+            or ""
+        ).strip(),
+        "llm_api_key": str(
+            payload.get("api_key")
+            or payload.get("llm_api_key")
+            or base_settings.get("llm_api_key")
+            or ""
+        ),
+        "llm_model": str(
+            payload.get("model")
+            or payload.get("llm_model")
+            or base_settings.get("llm_model")
+            or ""
+        ),
+        "llm_prompt": payload.get("prompt") or payload.get("llm_prompt") or base_settings.get("llm_prompt"),
+        "llm_context_mode": payload.get("context_mode") or base_settings.get("llm_context_mode"),
+        "llm_timeout": _coerce_float(payload.get("timeout"), base_settings.get("llm_timeout", 30.0)),
+        "normalization_apostrophe_mode": "llm",
+    }
+
+    merged = apply_normalization_overrides(base_settings, overrides)
+    if not merged.get("llm_base_url"):
+        return jsonify({"error": "LLM base URL is required."}), 400
+    if not merged.get("llm_model"):
+        return jsonify({"error": "Select an LLM model before previewing."}), 400
+
+    apostrophe_config = build_apostrophe_config(settings=merged)
+    try:
+        normalized_text = normalize_for_pipeline(sample_text, config=apostrophe_config, settings=merged)
+    except LLMClientError as exc:
+        return jsonify({"error": str(exc)}), 400
+
+    context = {
+        "text": sample_text,
+        "normalized_text": normalized_text,
+    }
+    return jsonify(context)
+
+
+@api_bp.post("/normalization/preview")
+def api_normalization_preview() -> ResponseReturnValue:
+    payload = request.get_json(force=True, silent=False) or {}
+    sample_text = str(payload.get("text") or "").strip()
+    if not sample_text:
+        return jsonify({"error": "Sample text is required."}), 400
+
+    base_settings = get_runtime_settings()
+    normalization_payload = payload.get("normalization") or {}
+    overrides: Dict[str, Any] = {}
+
+    boolean_keys = (
+        "normalization_numbers",
+        "normalization_titles",
+        "normalization_terminal",
+        "normalization_phoneme_hints",
+    )
+    for key in boolean_keys:
+        if key in normalization_payload:
+            overrides[key] = _coerce_bool(normalization_payload.get(key), base_settings.get(key, True))
+    if "normalization_apostrophe_mode" in normalization_payload:
+        overrides["normalization_apostrophe_mode"] = normalization_payload.get("normalization_apostrophe_mode")
+
+    llm_payload = payload.get("llm") or {}
+    for field in ("llm_base_url", "llm_api_key", "llm_model", "llm_prompt", "llm_context_mode"):
+        if field in llm_payload:
+            overrides[field] = llm_payload[field]
+    if "llm_timeout" in llm_payload:
+        overrides["llm_timeout"] = llm_payload.get("llm_timeout")
+
+    merged = apply_normalization_overrides(base_settings, overrides)
+
+    apostrophe_config = build_apostrophe_config(settings=merged)
+    try:
+        normalized_text = normalize_for_pipeline(sample_text, config=apostrophe_config, settings=merged)
+    except LLMClientError as exc:
+        return jsonify({"error": str(exc)}), 400
+
+    voice_spec = str(payload.get("voice") or base_settings.get("default_voice") or "").strip()
+    if not voice_spec and VOICES_INTERNAL:
+        voice_spec = VOICES_INTERNAL[0]
+    language = str(payload.get("language") or base_settings.get("language") or "a").strip() or "a"
+    try:
+        speed = float(payload.get("speed", 1.0) or 1.0)
+    except (TypeError, ValueError):
+        speed = 1.0
+    try:
+        max_seconds = max(1.0, min(15.0, float(payload.get("max_seconds", 8.0) or 8.0)))
+    except (TypeError, ValueError):
+        max_seconds = 8.0
+
+    use_gpu_default = base_settings.get("use_gpu", True)
+    use_gpu = _coerce_bool(payload.get("use_gpu"), use_gpu_default)
+
+    try:
+        audio_data = _synthesize_audio_from_normalized(
+            normalized_text=normalized_text,
+            voice_spec=voice_spec,
+            language=language,
+            speed=speed,
+            use_gpu=use_gpu,
+            max_seconds=max_seconds,
+        )
+    except ValueError as exc:
+        return jsonify({"error": str(exc)}), 400
+    except RuntimeError as exc:
+        return jsonify({"error": str(exc)}), 500
+
+    buffer = io.BytesIO()
+    sf.write(buffer, audio_data, SAMPLE_RATE, format="WAV")
+    audio_base64 = base64.b64encode(buffer.getvalue()).decode("ascii")
+
+    return jsonify(
+        {
+            "normalized_text": normalized_text,
+            "audio_base64": audio_base64,
+            "sample_rate": SAMPLE_RATE,
+        }
+    )
+
+
 @web_bp.get("/voices")
 def voice_profiles_page() -> str:
     options = _template_options()
@@ -2367,6 +2687,69 @@ def entities_override_update() -> ResponseReturnValue:
     return redirect(url_for("web.entities_page", **redirect_params))
 
 
+@api_bp.post("/entities/preview")
+def api_entity_pronunciation_preview() -> ResponseReturnValue:
+    payload = request.get_json(force=True, silent=False) or {}
+    token = str(payload.get("token") or "").strip()
+    pronunciation = str(payload.get("pronunciation") or "").strip()
+    if not token and not pronunciation:
+        return jsonify({"error": "Provide a token or pronunciation to preview."}), 400
+
+    settings = _load_settings()
+    sample_template = settings.get("speaker_pronunciation_sentence", "This is {{name}} speaking.")
+    spoken_label = pronunciation or token or ""
+    preview_text = _render_prompt_template(sample_template, {"name": spoken_label, "token": token})
+    if not preview_text.strip():
+        preview_text = spoken_label or token
+    if not preview_text:
+        return jsonify({"error": "Unable to construct preview text."}), 400
+
+    runtime_settings = get_runtime_settings()
+    apostrophe_config = build_apostrophe_config(settings=runtime_settings)
+    try:
+        normalized_text = normalize_for_pipeline(preview_text, config=apostrophe_config, settings=runtime_settings)
+    except LLMClientError as exc:
+        return jsonify({"error": str(exc)}), 400
+
+    voice_spec = str(payload.get("voice") or settings.get("default_voice") or "").strip()
+    if not voice_spec and VOICES_INTERNAL:
+        voice_spec = VOICES_INTERNAL[0]
+
+    language = str(payload.get("language") or runtime_settings.get("language") or "a").strip() or "a"
+    use_gpu = runtime_settings.get("use_gpu", True)
+    max_seconds = 6.0
+    try:
+        preview_speed = float(payload.get("speed", 1.0) or 1.0)
+    except (TypeError, ValueError):
+        preview_speed = 1.0
+    try:
+        audio_data = _synthesize_audio_from_normalized(
+            normalized_text=normalized_text,
+            voice_spec=voice_spec,
+            language=language,
+            speed=preview_speed,
+            use_gpu=use_gpu,
+            max_seconds=max_seconds,
+        )
+    except ValueError as exc:
+        return jsonify({"error": str(exc)}), 400
+    except RuntimeError as exc:
+        return jsonify({"error": str(exc)}), 500
+
+    buffer = io.BytesIO()
+    sf.write(buffer, audio_data, SAMPLE_RATE, format="WAV")
+    audio_base64 = base64.b64encode(buffer.getvalue()).decode("ascii")
+
+    return jsonify(
+        {
+            "text": preview_text,
+            "normalized_text": normalized_text,
+            "audio_base64": audio_base64,
+            "sample_rate": SAMPLE_RATE,
+        }
+    )
+
+
 @web_bp.route("/speakers", methods=["GET", "POST"])
 def speaker_configs_page() -> ResponseReturnValue:
     options = _template_options()

abogen/web/static/settings.js

@@ -0,0 +1,382 @@
+const form = document.querySelector('.settings__form');
+const navButtons = Array.from(document.querySelectorAll('.settings-nav__item'));
+const panels = Array.from(document.querySelectorAll('.settings-panel'));
+const llmNavButton = navButtons.find((button) => button.dataset.section === 'llm');
+
+const statusSelectors = {
+  llm: document.querySelector('[data-role="llm-preview-status"]'),
+  normalization: document.querySelector('[data-role="normalization-preview-status"]'),
+};
+
+const outputAreas = {
+  llm: document.querySelector('[data-role="llm-preview-output"]'),
+  normalization: document.querySelector('[data-role="normalization-preview-output"]'),
+};
+
+const normalizationAudio = document.querySelector('[data-role="normalization-preview-audio"]');
+
+function setStatus(target, message, state) {
+  if (!target) {
+    return;
+  }
+  target.textContent = message || '';
+  if (state) {
+    target.dataset.state = state;
+  } else {
+    delete target.dataset.state;
+  }
+}
+
+function clearStatus(target) {
+  setStatus(target, '', null);
+}
+
+function activatePanel(section) {
+  if (!section) {
+    return;
+  }
+  navButtons.forEach((button) => {
+    const isActive = button.dataset.section === section;
+    button.classList.toggle('is-active', isActive);
+  });
+  let activePanel = null;
+  panels.forEach((panel) => {
+    const isActive = panel.dataset.section === section;
+    panel.classList.toggle('is-active', isActive);
+    if (isActive) {
+      activePanel = panel;
+    }
+  });
+  if (activePanel) {
+    const focusable = activePanel.querySelector('input, select, textarea');
+    if (focusable) {
+      window.requestAnimationFrame(() => {
+        focusable.focus({ preventScroll: false });
+      });
+    }
+  }
+}
+
+function initNavigation() {
+  if (!navButtons.length || !panels.length) {
+    return;
+  }
+  navButtons.forEach((button) => {
+    button.addEventListener('click', () => {
+      activatePanel(button.dataset.section);
+      if (button.dataset.section) {
+        window.history.replaceState(null, '', `#${button.dataset.section}`);
+      }
+    });
+  });
+  const hash = window.location.hash.replace('#', '');
+  if (hash && panels.some((panel) => panel.dataset.section === hash)) {
+    activatePanel(hash);
+  } else {
+    const current = navButtons.find((button) => button.classList.contains('is-active'));
+    if (current) {
+      activatePanel(current.dataset.section);
+    }
+  }
+  window.addEventListener('hashchange', () => {
+    const section = window.location.hash.replace('#', '');
+    if (section) {
+      activatePanel(section);
+    }
+  });
+}
+
+function parseNumber(value, fallback) {
+  const parsed = Number.parseFloat(value);
+  return Number.isFinite(parsed) ? parsed : fallback;
+}
+
+function collectLLMFields() {
+  const baseUrl = form.querySelector('#llm_base_url');
+  const apiKey = form.querySelector('#llm_api_key');
+  const model = form.querySelector('#llm_model');
+  const prompt = form.querySelector('#llm_prompt');
+  const timeout = form.querySelector('#llm_timeout');
+  const context = form.querySelector('input[name="llm_context_mode"]:checked');
+  return {
+    base_url: baseUrl ? baseUrl.value.trim() : '',
+    api_key: apiKey ? apiKey.value.trim() : '',
+    model: model ? model.value.trim() : '',
+    prompt: prompt ? prompt.value : '',
+    context_mode: context ? context.value : 'sentence',
+    timeout: timeout ? parseNumber(timeout.value, 30) : 30,
+  };
+}
+
+function updateModelOptions(models) {
+  const select = form.querySelector('#llm_model');
+  if (!select) {
+    return;
+  }
+  const current = select.dataset.currentModel || select.value;
+  select.innerHTML = '';
+  if (!Array.isArray(models) || !models.length) {
+    const option = document.createElement('option');
+    option.value = '';
+    option.textContent = 'No models found';
+    select.appendChild(option);
+    select.dataset.currentModel = '';
+    select.disabled = true;
+    return;
+  }
+  const fragment = document.createDocumentFragment();
+  models.forEach((modelName) => {
+    const option = document.createElement('option');
+    option.value = modelName;
+    option.textContent = modelName;
+    if (modelName === current) {
+      option.selected = true;
+    }
+    fragment.appendChild(option);
+  });
+  select.appendChild(fragment);
+  select.dataset.currentModel = select.value || '';
+  select.disabled = false;
+}
+
+async function refreshModels(button) {
+  const status = statusSelectors.llm;
+  const llmFields = collectLLMFields();
+  if (!llmFields.base_url) {
+    setStatus(status, 'Enter a base URL before refreshing models.', 'error');
+    return;
+  }
+  clearStatus(status);
+  setStatus(status, 'Fetching models…');
+  button.disabled = true;
+  try {
+    const response = await fetch('/api/llm/models', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({
+        base_url: llmFields.base_url,
+        api_key: llmFields.api_key,
+        timeout: llmFields.timeout,
+      }),
+    });
+    const payload = await response.json();
+    if (!response.ok) {
+      throw new Error(payload.error || 'Unable to load models.');
+    }
+    updateModelOptions(payload.models || []);
+    const count = Array.isArray(payload.models) ? payload.models.length : 0;
+    if (count) {
+      setStatus(status, `Loaded ${count} model${count === 1 ? '' : 's'}.`, 'success');
+    } else {
+      setStatus(status, 'No models were returned.', 'error');
+    }
+  } catch (error) {
+    setStatus(status, error instanceof Error ? error.message : 'Failed to load models.', 'error');
+  } finally {
+    button.disabled = false;
+  }
+}
+
+async function previewLLM(button) {
+  const status = statusSelectors.llm;
+  const output = outputAreas.llm;
+  const previewText = document.querySelector('#llm_preview_text');
+  if (!previewText) {
+    return;
+  }
+  const llmFields = collectLLMFields();
+  if (!llmFields.base_url) {
+    setStatus(status, 'Enter a base URL to preview.', 'error');
+    return;
+  }
+  if (!llmFields.model) {
+    setStatus(status, 'Select a model to preview.', 'error');
+    return;
+  }
+  const sample = previewText.value.trim();
+  if (!sample) {
+    setStatus(status, 'Add some sample text first.', 'error');
+    return;
+  }
+  clearStatus(status);
+  if (output) {
+    output.textContent = '';
+  }
+  setStatus(status, 'Generating preview…');
+  button.disabled = true;
+  try {
+    const response = await fetch('/api/llm/preview', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({
+        text: sample,
+        base_url: llmFields.base_url,
+        api_key: llmFields.api_key,
+        model: llmFields.model,
+        prompt: llmFields.prompt,
+        context_mode: llmFields.context_mode,
+        timeout: llmFields.timeout,
+      }),
+    });
+    const payload = await response.json();
+    if (!response.ok) {
+      throw new Error(payload.error || 'Preview failed.');
+    }
+    if (output) {
+      output.textContent = payload.normalized_text || '';
+    }
+    setStatus(status, 'Preview ready.', 'success');
+  } catch (error) {
+    if (output) {
+      output.textContent = '';
+    }
+    setStatus(status, error instanceof Error ? error.message : 'Preview failed.', 'error');
+  } finally {
+    button.disabled = false;
+  }
+}
+
+function collectNormalizationSettings() {
+  const normalization = {
+    normalization_numbers: Boolean(form.querySelector('input[name="normalization_numbers"]')?.checked),
+    normalization_titles: Boolean(form.querySelector('input[name="normalization_titles"]')?.checked),
+    normalization_terminal: Boolean(form.querySelector('input[name="normalization_terminal"]')?.checked),
+    normalization_phoneme_hints: Boolean(form.querySelector('input[name="normalization_phoneme_hints"]')?.checked),
+    normalization_apostrophe_mode: form.querySelector('input[name="normalization_apostrophe_mode"]:checked')?.value || 'spacy',
+  };
+  return normalization;
+}
+
+function updateLLMNavState() {
+  if (!llmNavButton) {
+    return;
+  }
+  const fields = collectLLMFields();
+  if (fields.base_url && fields.api_key) {
+    llmNavButton.classList.remove('is-disabled');
+  } else {
+    llmNavButton.classList.add('is-disabled');
+  }
+}
+
+async function previewNormalization(button) {
+  const status = statusSelectors.normalization;
+  const output = outputAreas.normalization;
+  const textArea = document.querySelector('#normalization_sample_text');
+  const voiceSelect = document.querySelector('#normalization_sample_voice');
+  if (!textArea) {
+    return;
+  }
+  const sample = textArea.value.trim();
+  if (!sample) {
+    setStatus(status, 'Enter some text to preview.', 'error');
+    return;
+  }
+  clearStatus(status);
+  if (output) {
+    output.textContent = '';
+  }
+  if (normalizationAudio) {
+    normalizationAudio.hidden = true;
+    normalizationAudio.removeAttribute('src');
+  }
+  setStatus(status, 'Building preview…');
+  button.disabled = true;
+  try {
+    const normalization = collectNormalizationSettings();
+    const llmFields = collectLLMFields();
+    const response = await fetch('/api/normalization/preview', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({
+        text: sample,
+        voice: voiceSelect ? voiceSelect.value : undefined,
+        normalization,
+        llm: {
+          llm_base_url: llmFields.base_url,
+          llm_api_key: llmFields.api_key,
+          llm_model: llmFields.model,
+          llm_prompt: llmFields.prompt,
+          llm_context_mode: llmFields.context_mode,
+          llm_timeout: llmFields.timeout,
+        },
+        max_seconds: 8,
+      }),
+    });
+    const payload = await response.json();
+    if (!response.ok) {
+      throw new Error(payload.error || 'Preview failed.');
+    }
+    if (output) {
+      output.textContent = payload.normalized_text || '';
+    }
+    if (payload.audio_base64 && normalizationAudio) {
+      normalizationAudio.src = `data:audio/wav;base64,${payload.audio_base64}`;
+      normalizationAudio.hidden = false;
+      normalizationAudio.load();
+      normalizationAudio.play().catch(() => {
+        /* autoplay can fail; ignore */
+      });
+    }
+    setStatus(status, 'Preview updated.', 'success');
+  } catch (error) {
+    if (output) {
+      output.textContent = '';
+    }
+    if (normalizationAudio) {
+      normalizationAudio.hidden = true;
+      normalizationAudio.removeAttribute('src');
+    }
+    setStatus(status, error instanceof Error ? error.message : 'Preview failed.', 'error');
+  } finally {
+    button.disabled = false;
+  }
+}
+
+function initSampleSelector() {
+  const select = document.querySelector('#normalization_sample_select');
+  const textArea = document.querySelector('#normalization_sample_text');
+  if (!select || !textArea) {
+    return;
+  }
+  select.addEventListener('change', () => {
+    const option = select.selectedOptions[0];
+    if (option) {
+      textArea.value = option.value;
+    }
+  });
+}
+
+function initActions() {
+  const refreshButton = document.querySelector('[data-action="llm-refresh-models"]');
+  if (refreshButton) {
+    refreshButton.addEventListener('click', () => refreshModels(refreshButton));
+  }
+  const llmPreviewButton = document.querySelector('[data-action="llm-preview"]');
+  if (llmPreviewButton) {
+    llmPreviewButton.addEventListener('click', () => previewLLM(llmPreviewButton));
+  }
+  const normalizationButton = document.querySelector('[data-action="normalization-preview"]');
+  if (normalizationButton) {
+    normalizationButton.addEventListener('click', () => previewNormalization(normalizationButton));
+  }
+}
+
+function initLLMStateWatchers() {
+  const baseUrlInput = form.querySelector('#llm_base_url');
+  const apiKeyInput = form.querySelector('#llm_api_key');
+  if (!baseUrlInput || !apiKeyInput) {
+    return;
+  }
+  const handler = () => updateLLMNavState();
+  baseUrlInput.addEventListener('input', handler);
+  apiKeyInput.addEventListener('input', handler);
+  updateLLMNavState();
+}
+
+if (form) {
+  initNavigation();
+  initSampleSelector();
+  initActions();
+  initLLMStateWatchers();
+}

abogen/web/static/styles.css

@@ -1496,11 +1496,74 @@ button.step-indicator__item:focus-visible {
   box-shadow: 0 12px 30px rgba(239, 68, 68, 0.2);
 }
 
+
+.settings-layout {
+  display: grid;
+  grid-template-columns: 220px minmax(0, 1fr);
+  gap: 2rem;
+  align-items: start;
+}
+
+.settings-nav {
+  display: flex;
+  flex-direction: column;
+  gap: 0.5rem;
+  position: sticky;
+  top: 100px;
+}
+
+.settings-nav__item {
+  display: inline-flex;
+  justify-content: flex-start;
+  align-items: center;
+  padding: 0.65rem 0.9rem;
+  border-radius: 14px;
+  border: 1px solid rgba(148, 163, 184, 0.18);
+  background: rgba(15, 23, 42, 0.55);
+  color: var(--muted);
+  font-size: 0.95rem;
+  cursor: pointer;
+  transition: border 0.2s ease, color 0.2s ease, background 0.2s ease, box-shadow 0.2s ease;
+}
+
+.settings-nav__item:hover,
+.settings-nav__item:focus-visible {
+  border-color: rgba(56, 189, 248, 0.5);
+  color: var(--accent);
+  box-shadow: 0 0 0 3px rgba(56, 189, 248, 0.1);
+}
+
+.settings-nav__item.is-active {
+  background: rgba(56, 189, 248, 0.18);
+  border-color: rgba(56, 189, 248, 0.28);
+  color: #fff;
+  box-shadow: 0 10px 20px rgba(56, 189, 248, 0.18);
+}
+
+.settings-nav__item.is-disabled {
+  opacity: 0.6;
+  cursor: pointer;
+}
+
 .settings__form {
   display: grid;
   gap: 1.75rem;
 }
 
+.settings-panels {
+  display: grid;
+}
+
+.settings-panel {
+  display: none;
+  gap: 1.75rem;
+}
+
+.settings-panel.is-active {
+  display: grid;
+  gap: 1.75rem;
+}
+
 .settings__section {
   border: 1px solid rgba(148, 163, 184, 0.2);
   border-radius: 18px;
@@ -1586,6 +1649,140 @@ button.step-indicator__item:focus-visible {
   box-shadow: 0 0 0 4px rgba(56, 189, 248, 0.18);
 }
 
+.choices {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 0.5rem;
+}
+
+.choices--inline {
+  display: inline-flex;
+  flex-wrap: wrap;
+  gap: 0.5rem;
+}
+
+.radio-pill {
+  position: relative;
+  display: inline-flex;
+  align-items: center;
+  cursor: pointer;
+}
+
+.radio-pill input {
+  position: absolute;
+  opacity: 0;
+  inset: 0;
+  cursor: pointer;
+}
+
+.radio-pill span {
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  padding: 0.55rem 0.9rem;
+  border-radius: 999px;
+  border: 1px solid rgba(148, 163, 184, 0.25);
+  background: rgba(15, 23, 42, 0.4);
+  color: var(--muted);
+  transition: border 0.2s ease, background 0.2s ease, color 0.2s ease, box-shadow 0.2s ease;
+}
+
+.radio-pill:hover span {
+  border-color: rgba(56, 189, 248, 0.5);
+  color: var(--accent);
+}
+
+.radio-pill input:checked + span {
+  border-color: rgba(56, 189, 248, 0.45);
+  background: rgba(56, 189, 248, 0.16);
+  color: #fff;
+  box-shadow: 0 8px 18px rgba(56, 189, 248, 0.18);
+}
+
+.radio-pill input:focus-visible + span {
+  box-shadow: 0 0 0 4px rgba(56, 189, 248, 0.18);
+}
+
+.field__group {
+  display: flex;
+  flex-direction: column;
+  gap: 0.35rem;
+}
+
+.preview-card {
+  border: 1px solid rgba(148, 163, 184, 0.18);
+  border-radius: 16px;
+  padding: 1rem 1.1rem;
+  background: rgba(15, 23, 42, 0.52);
+  display: grid;
+  gap: 0.75rem;
+}
+
+.preview-card__actions {
+  display: flex;
+  align-items: center;
+  gap: 0.85rem;
+  flex-wrap: wrap;
+}
+
+.preview-card__status {
+  font-size: 0.85rem;
+  color: var(--muted);
+}
+
+.preview-card__status[data-state="error"] {
+  color: var(--danger);
+}
+
+.preview-card__status[data-state="success"] {
+  color: var(--success);
+}
+
+.preview-card__output {
+  border-radius: 12px;
+  background: rgba(15, 23, 42, 0.7);
+  border: 1px solid rgba(148, 163, 184, 0.2);
+  padding: 0.9rem 1rem;
+  font-family: "JetBrains Mono", "Fira Code", ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, "Liberation Mono", "Courier New", monospace;
+  font-size: 0.9rem;
+  line-height: 1.5;
+  color: var(--text);
+  min-height: 3.5rem;
+  white-space: pre-wrap;
+}
+
+.preview-card__output:empty {
+  display: none;
+}
+
+.preview-card__audio {
+  width: 100%;
+  margin-top: 0.25rem;
+}
+
+.hint--warning {
+  color: var(--warning);
+}
+
+@media (max-width: 860px) {
+  .settings-layout {
+    grid-template-columns: 1fr;
+  }
+
+  .settings-nav {
+    position: static;
+    flex-direction: row;
+    flex-wrap: wrap;
+    gap: 0.65rem;
+  }
+
+  .settings-nav__item {
+    flex: 1 1 calc(50% - 0.65rem);
+    justify-content: center;
+    text-align: center;
+  }
+}
+
 .prepare-summary {
   display: grid;
   grid-template-columns: minmax(0, 320px) minmax(0, 1fr);
@@ -1828,9 +2025,10 @@ button.step-indicator__item:focus-visible {
 }
 
 .field--inline {
-  display: inline-flex;
+  display: flex;
-  flex-direction: column;
+  flex-wrap: wrap;
-  gap: 0.35rem;
+  gap: 0.85rem;
+  align-items: flex-end;
 }
 
 .entity-summary__titles h2 {

abogen/web/templates/settings.html

@@ -11,7 +11,18 @@
   <div class="alert alert--success">Settings saved successfully.</div>
   {% endif %}
 
+  <div class="settings-layout">
+    <nav class="settings-nav" aria-label="Settings sections">
+      <button type="button" class="settings-nav__item is-active" data-section="narration">Narration</button>
+      <button type="button" class="settings-nav__item" data-section="audio">Audio &amp; Delivery</button>
+      <button type="button" class="settings-nav__item" data-section="subtitles">Subtitles &amp; Text</button>
+      <button type="button" class="settings-nav__item" data-section="performance">Performance</button>
+      <button type="button" class="settings-nav__item{% if not llm_ready %} is-disabled{% endif %}" data-section="llm">LLM</button>
+      <button type="button" class="settings-nav__item" data-section="normalization">Text Normalization</button>
+    </nav>
     <form action="{{ url_for('web.settings_page') }}" method="post" class="settings__form">
+      <div class="settings-panels">
+        <section class="settings-panel is-active" data-section="narration">
           <fieldset class="settings__section">
             <legend>Narration Defaults</legend>
             <div class="field">
@@ -59,7 +70,9 @@
               <p class="hint">Limits random voice selection for speakers marked as random. Leave empty to allow any language.</p>
             </div>
           </fieldset>
+        </section>
 
+        <section class="settings-panel" data-section="audio">
           <fieldset class="settings__section">
             <legend>Audio &amp; Delivery</legend>
             <div class="field">
@@ -118,7 +131,9 @@
               <p class="hint">Ensures the spoken chapter heading starts with "Chapter" when source titles begin with only a number or numeral.</p>
             </div>
           </fieldset>
+        </section>
 
+        <section class="settings-panel" data-section="subtitles">
           <fieldset class="settings__section">
             <legend>Subtitles &amp; Text</legend>
             <div class="field">
@@ -144,7 +159,9 @@
               </label>
             </div>
           </fieldset>
+        </section>
 
+        <section class="settings-panel" data-section="performance">
           <fieldset class="settings__section">
             <legend>Performance</legend>
             <div class="field field--choices">
@@ -154,10 +171,150 @@
               </label>
             </div>
           </fieldset>
+        </section>
+
+        <section class="settings-panel" data-section="llm">
+          <fieldset class="settings__section">
+            <legend>Endpoint</legend>
+            <div class="field">
+              <label for="llm_base_url">Base URL</label>
+              <input type="url" id="llm_base_url" name="llm_base_url" value="{{ settings.llm_base_url }}" placeholder="https://localhost:11434/v1">
+              <p class="hint">Point to an OpenAI-compatible endpoint such as Ollama or a proxy.</p>
+            </div>
+            <div class="field">
+              <label for="llm_api_key">API Key</label>
+              <input type="text" id="llm_api_key" name="llm_api_key" value="{{ settings.llm_api_key }}" autocomplete="off" placeholder="ollama">
+              <p class="hint">Leave blank or use <code>ollama</code> for local servers that do not require keys.</p>
+            </div>
+            <div class="field field--inline">
+              <div class="field__group">
+                <label for="llm_model">Default Model</label>
+                <select id="llm_model" name="llm_model" data-current-model="{{ settings.llm_model }}">
+                  {% if settings.llm_model %}
+                  <option value="{{ settings.llm_model }}" selected>{{ settings.llm_model }}</option>
+                  {% else %}
+                  <option value="" selected disabled>Select a model</option>
+                  {% endif %}
+                </select>
+              </div>
+              <div class="field__group">
+                <label for="llm_timeout">Timeout (seconds)</label>
+                <input type="number" step="1" min="1" id="llm_timeout" name="llm_timeout" value="{{ settings.llm_timeout }}">
+              </div>
+              <button type="button" class="button button--ghost" data-action="llm-refresh-models">Refresh models</button>
+            </div>
+          </fieldset>
+          <fieldset class="settings__section">
+            <legend>Normalization Prompt</legend>
+            <div class="field">
+              <label for="llm_prompt">Prompt Template</label>
+              <textarea id="llm_prompt" name="llm_prompt" rows="6">{{ settings.llm_prompt }}</textarea>
+              <p class="hint">Use placeholders like <code>{{ '{{sentence}}' }}</code> and <code>{{ '{{paragraph}}' }}</code> to inject content.</p>
+            </div>
+            <div class="field field--choices">
+              <span class="field__label">Context Mode</span>
+              <div class="choices choices--inline">
+                {% for option in llm_context_options %}
+                <label class="radio-pill">
+                  <input type="radio" name="llm_context_mode" value="{{ option.value }}" {% if settings.llm_context_mode == option.value %}checked{% endif %}>
+                  <span>{{ option.label }}</span>
+                </label>
+                {% endfor %}
+              </div>
+            </div>
+            <div class="preview-card" data-preview="llm">
+              <label for="llm_preview_text">Try the prompt</label>
+              <textarea id="llm_preview_text" rows="3">I've been waiting all day.</textarea>
+              <div class="preview-card__actions">
+                <button type="button" class="button" data-action="llm-preview">Preview</button>
+                <span class="preview-card__status" data-role="llm-preview-status"></span>
+              </div>
+              <div class="preview-card__output" data-role="llm-preview-output"></div>
+            </div>
+          </fieldset>
+        </section>
+
+        <section class="settings-panel" data-section="normalization">
+          <fieldset class="settings__section">
+            <legend>Normalization Rules</legend>
+            <div class="field field--choices">
+              <label class="toggle-pill">
+                <input type="checkbox" name="normalization_numbers" value="true" {% if settings.normalization_numbers %}checked{% endif %}>
+                <span>Convert grouped numbers to words</span>
+              </label>
+              <label class="toggle-pill">
+                <input type="checkbox" name="normalization_titles" value="true" {% if settings.normalization_titles %}checked{% endif %}>
+                <span>Expand titles and suffixes (Dr., St., Jr., …)</span>
+              </label>
+              <label class="toggle-pill">
+                <input type="checkbox" name="normalization_terminal" value="true" {% if settings.normalization_terminal %}checked{% endif %}>
+                <span>Ensure sentences end with terminal punctuation</span>
+              </label>
+              <label class="toggle-pill">
+                <input type="checkbox" name="normalization_phoneme_hints" value="true" {% if settings.normalization_phoneme_hints %}checked{% endif %}>
+                <span>Add phoneme hints for possessives</span>
+              </label>
+            </div>
+            <div class="field">
+              <span class="field__label">Apostrophe strategy</span>
+              <div class="choices choices--inline">
+                {% for option in apostrophe_modes %}
+                <label class="radio-pill">
+                  <input type="radio" name="normalization_apostrophe_mode" value="{{ option.value }}" {% if settings.normalization_apostrophe_mode == option.value %}checked{% endif %}>
+                  <span>{{ option.label }}</span>
+                </label>
+                {% endfor %}
+              </div>
+              {% if settings.normalization_apostrophe_mode == 'llm' and not llm_ready %}
+              <p class="hint hint--warning">Configure the LLM connection before using it for audiobook runs.</p>
+              {% endif %}
+            </div>
+          </fieldset>
+          <fieldset class="settings__section">
+            <legend>Sample &amp; Preview</legend>
+            <div class="field field--inline">
+              <div class="field__group">
+                <label for="normalization_sample_select">Sample</label>
+                <select id="normalization_sample_select">
+                  {% for key, text in normalization_samples.items() %}
+                  <option value="{{ text }}" {% if loop.first %}selected{% endif %}>{{ key|capitalize }}</option>
+                  {% endfor %}
+                </select>
+              </div>
+              <div class="field__group">
+                <label for="normalization_sample_voice">Voice</label>
+                <select id="normalization_sample_voice">
+                  {% for voice in options.voices %}
+                  <option value="{{ voice }}" {% if settings.default_voice == voice %}selected{% endif %}>{{ voice }}</option>
+                  {% endfor %}
+                </select>
+              </div>
+            </div>
+            <div class="field">
+              <label for="normalization_sample_text">Sample text</label>
+              <textarea id="normalization_sample_text" rows="4">{{ normalization_samples['apostrophes'] }}</textarea>
+            </div>
+            <div class="preview-card" data-preview="normalization">
+              <div class="preview-card__actions">
+                <button type="button" class="button" data-action="normalization-preview">Preview with current settings</button>
+                <span class="preview-card__status" data-role="normalization-preview-status"></span>
+              </div>
+              <pre class="preview-card__output" data-role="normalization-preview-output"></pre>
+              <audio controls class="preview-card__audio" data-role="normalization-preview-audio" hidden></audio>
+            </div>
+          </fieldset>
+        </section>
+      </div>
 
       <div class="settings__actions">
         <button type="submit" class="button">Save Settings</button>
       </div>
     </form>
+  </div>
 </section>
 {% endblock %}
+
+{% block scripts %}
+{{ super() }}
+<script type="module" src="{{ url_for('static', filename='settings.js') }}"></script>
+{% endblock %}