From 44e555c83e73c25446e320a73388106ff10e9b14 Mon Sep 17 00:00:00 2001
From: Juraj Borza <jborza@gmail.com>
Date: Mon, 28 Apr 2025 09:19:48 +0200
Subject: [PATCH] added voice formulas for mixed text

---
 abogen/conversion.py     |  6 +++++-
 abogen/gui.py            | 14 +++++++++++--
 abogen/voice_formulas.py | 43 ++++++++++++++++++++++++++++++++++++++++
 3 files changed, 60 insertions(+), 3 deletions(-)
 create mode 100644 abogen/voice_formulas.py

diff --git a/abogen/conversion.py b/abogen/conversion.py
index c2eadb3..66386dd 100644
--- a/abogen/conversion.py
+++ b/abogen/conversion.py
@@ -9,6 +9,7 @@ from PyQt5.QtWidgets import QCheckBox, QVBoxLayout, QDialog, QLabel, QDialogButt
 import soundfile as sf
 from utils import clean_text
 from constants import PROGRAM_NAME, LANGUAGE_DESCRIPTIONS, SAMPLE_VOICE_TEXTS
+from voice_formulas import get_new_voice
 
 
 def get_sample_voice_text(lang_code):
@@ -346,9 +347,12 @@ class ConversionThread(QThread):
 
                 # Set split_pattern to \n+ which will split on one or more newlines
                 split_pattern = r"\n+"
+                
+                loaded_voice = get_new_voice(tts, self.voice, self.use_gpu)
+                
                 for result in tts(
                     chapter_text,
-                    voice=self.voice,
+                    voice=loaded_voice,
                     speed=self.speed,
                     split_pattern=split_pattern,
                 ):
diff --git a/abogen/gui.py b/abogen/gui.py
index c8e30ed..269e12e 100644
--- a/abogen/gui.py
+++ b/abogen/gui.py
@@ -567,6 +567,13 @@ class abogen(QWidget):
             '"a" => American English\n"b" => British English\n"e" => Spanish\n"f" => French\n"h" => Hindi\n"i" => Italian\n"j" => Japanese\n"p" => Brazilian Portuguese\n"z" => Mandarin Chinese\nThe second character represents the gender:\n"m" => Male\n"f" => Female'
         )
         voice_row.addWidget(self.voice_combo)
+        
+        # voice formula - text box for now, add a dialog later
+        self.voice_formula = QTextEdit(self)
+        self.voice_formula.setAcceptRichText(False)
+        self.voice_formula.setPlaceholderText("Enter voice formula here...")
+        self.voice_formula.setText('0.242 * am_michael + 0.758 * bf_isabella')
+        voice_row.addWidget(self.voice_formula)
 
         # Play/Stop icons
         def make_icon(color, shape):
@@ -1037,12 +1044,15 @@ class abogen(QWidget):
                 if not self.subtitle_combo.isEnabled()
                 else self.subtitle_mode
             )
-
+            
+            voice_formula = self.voice_formula.toPlainText()
+            
             self.conversion_thread = ConversionThread(
                 self.selected_file,
                 self.selected_lang,
                 speed,
-                self.selected_voice,
+                #self.selected_voice,
+                voice_formula,
                 self.save_option,
                 self.selected_output_folder,
                 subtitle_mode=actual_subtitle_mode,
diff --git a/abogen/voice_formulas.py b/abogen/voice_formulas.py
new file mode 100644
index 0000000..29272e5
--- /dev/null
+++ b/abogen/voice_formulas.py
@@ -0,0 +1,43 @@
+from constants import VOICES_INTERNAL
+
+# Calls parsing and loads the voice to gpu or cpu
+def get_new_voice(pipeline, formula, use_gpu):
+    try:
+        weighted_voice = parse_voice_formula(pipeline, formula)            
+        device = "cuda" if use_gpu else "cpu"
+        return weighted_voice.to(device)
+    except Exception as e:
+        raise ValueError(f"Failed to create voice: {str(e)}")
+    
+# Parse the formula and get the combined voice tensor        
+def parse_voice_formula(pipeline, formula):
+    """Parse the voice formula string and return the combined voice tensor."""
+    if not formula.strip():
+        raise ValueError("Empty voice formula")
+        
+    # Initialize the weighted sum
+    weighted_sum = None
+    
+    # Split the formula into terms
+    terms = formula.split('+')
+    
+    for term in terms:
+        # Parse each term (format: "0.333 * voice_name")
+        weight, voice_name = term.strip().split('*')
+        weight = float(weight.strip())
+        voice_name = voice_name.strip()
+        
+        # Get the voice tensor
+        # use VOICES_INTERNAL
+        if voice_name not in VOICES_INTERNAL:
+            raise ValueError(f"Unknown voice: {voice_name}")
+            
+        voice_tensor = pipeline.load_single_voice(voice_name)
+        
+        # Add to weighted sum
+        if weighted_sum is None:
+            weighted_sum = weight * voice_tensor
+        else:
+            weighted_sum += weight * voice_tensor
+            
+    return weighted_sum
\ No newline at end of file