Update app.py

app.py

@@ -6,7 +6,7 @@ import numpy as np
 import wave
 
 #tts
-from balacoon_tts import TTS
+from balacoon_tts import TTS, SpeechUtterance
 from threading import Lock
 from huggingface_hub import hf_hub_download, list_repo_files
 from pydub import AudioSegment
@@ -100,16 +100,16 @@ async def greet(product,description):
     response = client.chat.completions.create(model=guard_llm, messages=messages, temperature=0)
     if response.choices[0].message.content != "not moderated":
         a_list = ["Sorry, I can't proceed for generating marketing email. Your content needs to be moderated first. Thank you!"]
-        with locker:
-            tts = TTS(os.path.join(os.getcwd(), tts_model_str))
-            speakers = tts.get_speakers()
-            if len(a_list[0]) > 1024:
-                # truncate the text
-                text_str = a_list[0][:1024]
-            else:
-                text_str = a_list[0]
-            samples = tts.synthesize(text_str, speakers[-1])
-        yield gr.Audio(value=(tts.get_sampling_rate(), samples))
+        tts = TTS(os.path.join(os.getcwd(), tts_model_str))
+        # create an utterance, which will hold the syntheis state
+        utterance = SpeechUtterance(a_list[0])
+        # run synthesis in a loop
+        while True:
+            samples = tts.synthesize_chunk(utterance, tts.get_speakers()[-1])
+            if len(samples) == 0:
+                # all the samples were already generated
+                break
+            yield gr.Audio(value=(tts.get_sampling_rate(), samples))
     else:
         output = llm.create_chat_completion(
         messages=[
@@ -124,14 +124,20 @@ async def greet(product,description):
         stream=True
         )
         partial_message = ""
+        tts = TTS(os.path.join(os.getcwd(), tts_model_str))
         for chunk in output:
             delta = chunk['choices'][0]['delta']
             if 'content' in delta:
-                with locker:
-                    tts = TTS(os.path.join(os.getcwd(), tts_model_str))
-                    speakers = tts.get_speakers()
-                    samples = tts.synthesize(delta.get('content', ''), speakers[-1])
-        yield (tts.get_sampling_rate(), samples)
+                partial_message = partial_message + delta.get('content', '')
+        # create an utterance, which will hold the syntheis state
+        utterance = SpeechUtterance(partial_message)
+        # run synthesis in a loop
+        while True:
+            samples = tts.synthesize_chunk(utterance, tts.get_speakers()[-1])
+            if len(samples) == 0:
+                # all the samples were already generated
+                break
+            yield gr.Audio(value=(tts.get_sampling_rate(), samples))
 
 demo = gr.Interface(fn=greet, inputs=["text","text"], outputs=gr.Audio(), concurrency_limit=10)
 demo.launch()