Update app.py

app.py

@@ -88,6 +88,8 @@ api = HfApi(token=HF_TOKEN)
 
 repo_id = "jbilcke-hf/zephyr-xtts"
 
+SENTENCE_SPLIT_LENGTH=250
+
 default_system_message = f"""
 You're the storyteller, crafting a short tale for young listeners. Please abide by these guidelines:
 - Keep your sentences short, concise and easy to understand.
@@ -143,6 +145,26 @@ LLAMA_VERBOSE=False
 print("Running LLM Zephyr")
 llm_zephyr = Llama(model_path=zephyr_model_path,n_gpu_layers=GPU_LAYERS-10,max_new_tokens=256, context_window=4096, n_ctx=4096,n_batch=128,verbose=LLAMA_VERBOSE)
 
+def split_sentences(text, max_len):
+    # Apply custom rules to enforce sentence breaks with double punctuation
+    text = re.sub(r"(\s*\.{2})\s*", r".\1 ", text)  # for '..'
+    text = re.sub(r"(\s*\!{2})\s*", r"!\1 ", text)  # for '!!'
+
+    # Use NLTK to split into sentences
+    sentences = nltk.sent_tokenize(text)
+
+    # Then check if each sentence is greater than max_len, if so, use textwrap to split it
+    sentence_list = []
+    for sent in sentences:
+        if len(sent) > max_len:
+            wrapped = textwrap.wrap(sent, max_len, break_long_words=True)
+            sentence_list.extend(wrapped)
+        else:
+            sentence_list.append(sent)
+    
+    return sentence_list
+
+
 # <|system|>
 # You are a friendly chatbot who always responds in the style of a pirate.</s>
 # <|user|>
@@ -507,7 +529,6 @@ def generate_speech_for_sentence(history, chatbot_role, sentence, return_as_byte
     print("Sentence for speech:", sentence)
 
     try:
-        SENTENCE_SPLIT_LENGTH=350
         if len(sentence)<SENTENCE_SPLIT_LENGTH:
             # no problem continue on
             sentence_list = [sentence]
@@ -515,7 +536,8 @@ def generate_speech_for_sentence(history, chatbot_role, sentence, return_as_byte
             # Until now nltk likely split sentences properly but we need additional 
             # check for longer sentence and split at last possible position
             # Do whatever necessary, first break at hypens then spaces and then even split very long words
-            sentence_list=textwrap.wrap(sentence,SENTENCE_SPLIT_LENGTH)
+            # sentence_list=textwrap.wrap(sentence,SENTENCE_SPLIT_LENGTH)
+            sentence_list = split_sentences(sentence, SENTENCE_SPLIT_LENGTH)
             print("SPLITTED LONG SENTENCE:",sentence_list)
         
         for sentence in sentence_list: