Merge pull request #325 from ROBERT-MCDOWELL/v25

fix splitting sentences blanks. various fixes

lib/classes/tts_manager.py

@@ -63,6 +63,7 @@ def __init__(self, session):
         self._build()
 
     def _build(self):
+        self.params['current_voice_path'] = None
         if self.session['tts_engine'] == XTTSv2:
             if self.session['custom_model'] is not None:
                 self.model_name = os.path.basename(self.session['custom_model'])
@@ -168,15 +169,17 @@ def convert_sentence_to_audio(self):
             '''
             if self.session['tts_engine'] == XTTSv2:
                 if self.session['custom_model'] is not None or self.session['fine_tuned'] != 'internal':
-                    msg = 'Computing speaker latents...'
-                    print(msg)
                     self.params['voice_path'] = (
                         self.session['voice'] if self.session['voice'] is not None 
                         else os.path.join(self.session['custom_model_dir'], self.session['tts_engine'], self.session['custom_model'],'ref.wav') if self.session['custom_model']
                         else models[self.session['tts_engine']][self.session['fine_tuned']]['voice'] if self.session['fine_tuned']
                         else models[self.session['tts_engine']]['internal']['voice']
                     )
-                    self.params['gpt_cond_latent'], self.params['speaker_embedding'] = self.params['tts'].get_conditioning_latents(audio_path=[self.params['voice_path']])
+                    if self.params['current_voice_path'] != self.params['voice_path']:
+                        msg = 'Computing speaker latents...'
+                        print(msg)
+                        self.params['current_voice_path'] = self.params['voice_path']
+                        self.params['gpt_cond_latent'], self.params['speaker_embedding'] = self.params['tts'].get_conditioning_latents(audio_path=[self.params['voice_path']])
                     with torch.no_grad():
                         result = self.params['tts'].inference(
                             text=self.params['sentence'],

lib/functions.py

@@ -549,7 +549,8 @@ def filter_chapter(doc, lang, lang_iso1, tts_engine):
     text = normalize_text(text, lang, lang_iso1, tts_engine)
     # Create regex pattern from punctuation list to split the phoneme_list
     escaped_punctuation = re.escape(''.join(punctuation_list))
-    punctuation_pattern_split = rf'([^{"".join(escaped_punctuation)}]+|[{escaped_punctuation}])'
+    #punctuation_pattern_split = rf'([^{"".join(escaped_punctuation)}]+|[{escaped_punctuation}])'
+    punctuation_pattern_split = rf'(\S.*?[{"".join(escaped_punctuation)}])|\S+'
     # Split by punctuation marks while keeping the punctuation at the end of each word
     tmp_list = re.findall(punctuation_pattern_split, text)
     phoneme_list =  [phoneme.strip() for phoneme in tmp_list if phoneme.strip()]    
@@ -575,7 +576,7 @@ def filter_pattern(doc_identifier):
         elif re.match(r'^\d+$', segment):
             return 'numbers'
     return None
-
+'''
 def get_sentences(phoneme_list, max_tokens):
     sentences = []
     current_sentence = ""
@@ -602,7 +603,42 @@ def get_sentences(phoneme_list, max_tokens):
     if current_sentence:
         sentences.append(current_sentence.strip())
     return sentences
-
+'''
+
+def get_sentences(phoneme_list, max_tokens):
+    sentences = []
+    current_sentence = ""
+    current_phoneme_count = 0
+    for phoneme in phoneme_list:
+        part_phoneme_count = len(phoneme.split())
+        # Always append to current sentence unless punctuation is hit
+        if current_phoneme_count + part_phoneme_count > max_tokens:
+            # Ensure we finalize the sentence at punctuation, not a space
+            if any(current_sentence.endswith(punc) for punc in punctuation_list):
+                sentences.append(current_sentence.strip())
+                current_sentence = phoneme
+                current_phoneme_count = part_phoneme_count
+            else:
+                # Look back and split at last punctuation instead of splitting randomly
+                last_punc_index = max(
+                    (current_sentence.rfind(punc) for punc in punctuation_list if punc in current_sentence),
+                    default=-1
+                )
+                if last_punc_index > -1:
+                    sentences.append(current_sentence[:last_punc_index+1].strip())  # Keep punctuation
+                    current_sentence = current_sentence[last_punc_index+1:].strip() + " " + phoneme
+                    current_phoneme_count = len(current_sentence.split())
+                else:
+                    sentences.append(current_sentence.strip())
+                    current_sentence = phoneme
+                    current_phoneme_count = part_phoneme_count
+        else:
+            current_sentence += (" " if current_sentence else "") + phoneme
+            current_phoneme_count += part_phoneme_count
+    if current_sentence:
+        sentences.append(current_sentence.strip())
+    return sentences
+
 def get_batch_size(list, session):
     total_size = 0
     print(list)
@@ -1596,7 +1632,6 @@ def process_cleanup(state):
                     label='Enable Text Splitting', 
                     value=default_xtts_settings['enable_text_splitting'],
                     info='Coqui-tts builtin text splitting. Can help against hallucinations bu can also be worse.',
-                    visible=False
                 )
 
         gr_state = gr.State(value={"hash": None})
@@ -2076,6 +2111,8 @@ def change_gr_fine_tuned_list(selected, id):
             visible = False
             if selected == 'internal' and session['tts_engine'] == XTTSv2:
                 visible = visible_gr_group_custom_model
+            else:
+                visible = False
             session['fine_tuned'] = selected
             return gr.update(visible=visible)