abogen/abogen/webui/routes/utils/synthesize.py

import io
import logging
import threading
from typing import Any, Dict, Iterable, List, Mapping, Optional, Tuple
import numpy as np
import soundfile as sf
from flask import current_app, send_file
from flask.typing import ResponseReturnValue

from abogen.domain.audio_helpers import to_float32
from abogen.domain.device import select_device as _select_device
from abogen.domain.enums import Language
from abogen.domain.split_pattern import get_split_pattern
from abogen.domain.pronunciation import (
    merge_pronunciation_overrides,
    compile_pronunciation_rules,
    apply_pronunciation_rules,
)


SAMPLE_RATE = 24000

_preview_pipelines: Dict[Tuple[str, str], Any] = {}
_preview_pipeline_lock = threading.Lock()


def clear_preview_pipelines() -> None:
    """Dispose all cached preview pipelines and clear the cache."""
    with _preview_pipeline_lock:
        for pipeline in _preview_pipelines.values():
            try:
                pipeline.dispose()
            except Exception:
                pass
        _preview_pipelines.clear()


def _resolve_pipeline(language: Language, use_gpu: bool) -> Tuple[Any, bool]:
    devices: List[str] = ["cpu"]
    if use_gpu:
        preferred = _select_device()
        if preferred != "cpu":
            devices.insert(0, preferred)

    last_error: Optional[Exception] = None
    for device in devices:
        try:
            logging.info("[preview] Trying device=%s for language=%s", device, language)
            return get_preview_pipeline(language, device), device != "cpu"
        except Exception as exc:
            last_error = exc
            logging.warning("[preview] Device %s failed: %s", device, exc)

    raise RuntimeError("Preview pipeline is unavailable") from last_error


def get_preview_pipeline(language: Language, device: str) -> Any:
    key = (language, device)
    with _preview_pipeline_lock:
        pipeline = _preview_pipelines.get(key)
        if pipeline is not None:
            logging.info("[preview] Using cached pipeline for %s/%s", language, device)
            return pipeline
        from abogen.tts_plugin.utils import create_pipeline

        logging.info("[preview] Creating pipeline: provider=kokoro language=%s device=%s", language, device)
        pipeline = create_pipeline("kokoro", language=language, device=device)
        _preview_pipelines[key] = pipeline
        return pipeline

def generate_preview_audio(
    text: str,
    voice_spec: str,
    language: Language,
    speed: float,
    use_gpu: bool,
    tts_provider: str = "kokoro",
    supertonic_total_steps: int = 5,
    max_seconds: float = 8.0,
    pronunciation_overrides: Optional[Iterable[Mapping[str, Any]]] = None,
    manual_overrides: Optional[Iterable[Mapping[str, Any]]] = None,
    speakers: Optional[Mapping[str, Any]] = None,
) -> bytes:
    if not text.strip():
        raise ValueError("Preview text is required")

    provider = (tts_provider or "kokoro").strip().lower()

    # Apply pronunciation/manual overrides first so tokens like `Unfu*k` still match
    # before any downstream normalization potentially strips punctuation.
    source_text = text
    if pronunciation_overrides or manual_overrides or speakers:
        try:
            class _PreviewJob:
                def __init__(self):
                    self.language = language
                    self.voice = voice_spec
                    self.speakers = speakers
                    self.manual_overrides = list(manual_overrides or [])
                    self.pronunciation_overrides = list(pronunciation_overrides or [])

            job = _PreviewJob()
            merged = merge_pronunciation_overrides(job)
            rules = compile_pronunciation_rules(merged)
            source_text = apply_pronunciation_rules(source_text, rules)
        except Exception:
            current_app.logger.exception("Preview override application failed; using raw text")
            source_text = text

    normalized_text = source_text
    if provider != "supertonic":
        try:
            from abogen.kokoro_text_normalization import normalize_for_pipeline

            normalized_text = normalize_for_pipeline(source_text)
        except Exception:
            current_app.logger.exception("Preview normalization failed; using raw text")
            normalized_text = source_text

    preview_split = get_split_pattern(language, "Disabled")

    if provider == "supertonic":
        from abogen.tts_plugin.utils import create_pipeline

        pipeline = create_pipeline("supertonic", language=language)
        segments = pipeline(
            normalized_text,
            voice=voice_spec,
            speed=speed,
            split_pattern=preview_split,
            total_steps=supertonic_total_steps,
        )
    else:
        pipeline, pipeline_uses_gpu = _resolve_pipeline(language, use_gpu)
        if pipeline is None:
            raise RuntimeError("Preview pipeline is unavailable")

        voice_choice: Any = voice_spec
        if voice_spec and "*" in voice_spec:
            from abogen.domain.voice_loader import resolve_voice

            voice_choice = resolve_voice(voice_spec, pipeline, pipeline_uses_gpu)

        segments = pipeline(
            normalized_text,
            voice=voice_choice,
            speed=speed,
            split_pattern=preview_split,
        )

    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")

    audio_data = np.concatenate(audio_chunks)
    buffer = io.BytesIO()
    sf.write(buffer, audio_data, SAMPLE_RATE, format="WAV")
    return buffer.getvalue()

def synthesize_preview(
    text: str,
    voice_spec: str,
    language: Language,
    speed: float,
    use_gpu: bool,
    tts_provider: str = "kokoro",
    supertonic_total_steps: int = 5,
    max_seconds: float = 8.0,
    pronunciation_overrides: Optional[Iterable[Mapping[str, Any]]] = None,
    manual_overrides: Optional[Iterable[Mapping[str, Any]]] = None,
    speakers: Optional[Mapping[str, Any]] = None,
) -> ResponseReturnValue:
    try:
        audio_bytes = generate_preview_audio(
            text=text,
            voice_spec=voice_spec,
            language=language,
            speed=speed,
            use_gpu=use_gpu,
            tts_provider=tts_provider,
            supertonic_total_steps=supertonic_total_steps,
            max_seconds=max_seconds,
            pronunciation_overrides=pronunciation_overrides,
            manual_overrides=manual_overrides,
            speakers=speakers,
        )
    except Exception as e:
        raise e

    buffer = io.BytesIO(audio_bytes)
    response = send_file(
        buffer,
        mimetype="audio/wav",
        as_attachment=False,
        download_name="speaker_preview.wav",
    )
    response.headers["Cache-Control"] = "no-store"
    return response