Update app.py

app.py

@@ -40,6 +40,9 @@ llm = Llama.from_pretrained(
     verbose=False
 )
 
+# locker that disallow access to the tts object from more then one thread
+locker = Lock()
+
 #guardrail model
 guard_llm = "llama-3.1-8b-instant"
 
@@ -95,14 +98,15 @@ 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!"]
-        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])
+        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))
     else:
         output = llm.create_chat_completion(
@@ -118,14 +122,15 @@ def greet(product,description):
         stream=True
         )
         partial_message = ""
-        tts = TTS(os.path.join(os.getcwd(), tts_model_str))
-        speakers = tts.get_speakers()
         for chunk in output:
             delta = chunk['choices'][0]['delta']
             if 'content' in delta:
-                samples = tts.synthesize(delta.get('content', ''), speakers[-1])
+                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 gr.Audio(value=(tts.get_sampling_rate(), samples))
 
-
-demo = gr.Interface(fn=greet, inputs=["text","text"], concurrency_limit=10)
+audio = gr.Audio()
+demo = gr.Interface(fn=greet, inputs=["text","text"], concurrency_limit=10, outputs=audio)
 demo.launch()