Update app.py

app.py

@@ -755,18 +755,87 @@ repo_id = "parler-tts/parler-tts-mini-v1"
 
 
 
+# def generate_audio_parler_tts(text):
+#     description = "A female speaker delivers a slightly expressive and animated speech with a moderate speed and pitch. The recording is of very high quality, with the speaker's voice sounding clear and very close up."
+#     chunk_size_in_s = 0.5
+
+#     # Initialize the tokenizer and model
+#     parler_tokenizer = AutoTokenizer.from_pretrained(repo_id)
+#     parler_model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id).to(device)
+#     sampling_rate = parler_model.audio_encoder.config.sampling_rate
+#     frame_rate = parler_model.audio_encoder.config.frame_rate
+
+#     def generate(text, description, play_steps_in_s=0.5):
+#         play_steps = int(frame_rate * play_steps_in_s)
+#         streamer = ParlerTTSStreamer(parler_model, device=device, play_steps=play_steps)
+
+#         inputs = parler_tokenizer(description, return_tensors="pt").to(device)
+#         prompt = parler_tokenizer(text, return_tensors="pt").to(device)
+
+#         generation_kwargs = dict(
+#             input_ids=inputs.input_ids,
+#             prompt_input_ids=prompt.input_ids,
+#             attention_mask=inputs.attention_mask,
+#             prompt_attention_mask=prompt.attention_mask,
+#             streamer=streamer,
+#             do_sample=True,
+#             temperature=1.0,
+#             min_new_tokens=10,
+#         )
+
+#         thread = Thread(target=parler_model.generate, kwargs=generation_kwargs)
+#         thread.start()
+
+#         for new_audio in streamer:
+#             if new_audio.shape[0] == 0:
+#                 break
+#             # Save or process each audio chunk as it is generated
+#             yield sampling_rate, new_audio
+
+#     audio_segments = []
+#     for (sampling_rate, audio_chunk) in generate(text, description, chunk_size_in_s):
+#         audio_segments.append(audio_chunk)
+#         # Here, you can save the chunk to a file or send it to a frontend
+#         # For example, you could write the chunk to a file immediately:
+#         temp_audio_path = os.path.join(tempfile.gettempdir(), f"parler_tts_audio_chunk_{len(audio_segments)}.wav")
+#         write_wav(temp_audio_path, sampling_rate, audio_chunk.astype(np.float32))
+#         logging.debug(f"Saved chunk to {temp_audio_path}")
+#         # You could also send the chunk to a web client if this was a web application
+
+#     # Combine all the audio chunks into one audio file
+#     combined_audio = np.concatenate(audio_segments)
+#     combined_audio_path = os.path.join(tempfile.gettempdir(), "parler_tts_combined_audio_stream.wav")
+
+#     write_wav(combined_audio_path, sampling_rate, combined_audio.astype(np.float32))
+
+#     logging.debug(f"Combined audio saved to {combined_audio_path}")
+#     return combined_audio_path
+
+import concurrent.futures
+import tempfile
+import os
+import numpy as np
+from threading import Thread
+from transformers import AutoTokenizer
+from parler_tts import ParlerTTSForConditionalGeneration, ParlerTTSStreamer
+from scipy.io.wavfile import write as write_wav
+import logging
+
 def generate_audio_parler_tts(text):
     description = "A female speaker delivers a slightly expressive and animated speech with a moderate speed and pitch. The recording is of very high quality, with the speaker's voice sounding clear and very close up."
-    chunk_size_in_s = 0.5
+    
+    chunk_size_in_s = 0.3  # Smaller chunk size for lower latency
 
     # Initialize the tokenizer and model
     parler_tokenizer = AutoTokenizer.from_pretrained(repo_id)
     parler_model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id).to(device)
+    
     sampling_rate = parler_model.audio_encoder.config.sampling_rate
     frame_rate = parler_model.audio_encoder.config.frame_rate
+    
+    play_steps = int(frame_rate * chunk_size_in_s)
 
-    def generate(text, description, play_steps_in_s=0.5):
-        play_steps = int(frame_rate * play_steps_in_s)
+    def generate_chunks(text, description):
         streamer = ParlerTTSStreamer(parler_model, device=device, play_steps=play_steps)
 
         inputs = parler_tokenizer(description, return_tensors="pt").to(device)
@@ -779,7 +848,7 @@ def generate_audio_parler_tts(text):
             prompt_attention_mask=prompt.attention_mask,
             streamer=streamer,
             do_sample=True,
-            temperature=1.0,
+            temperature=0.7,  # Lower temperature for faster generation
             min_new_tokens=10,
         )
 
@@ -789,18 +858,27 @@ def generate_audio_parler_tts(text):
         for new_audio in streamer:
             if new_audio.shape[0] == 0:
                 break
-            # Save or process each audio chunk as it is generated
             yield sampling_rate, new_audio
 
-    audio_segments = []
-    for (sampling_rate, audio_chunk) in generate(text, description, chunk_size_in_s):
-        audio_segments.append(audio_chunk)
-        # Here, you can save the chunk to a file or send it to a frontend
-        # For example, you could write the chunk to a file immediately:
-        temp_audio_path = os.path.join(tempfile.gettempdir(), f"parler_tts_audio_chunk_{len(audio_segments)}.wav")
-        write_wav(temp_audio_path, sampling_rate, audio_chunk.astype(np.float32))
-        logging.debug(f"Saved chunk to {temp_audio_path}")
-        # You could also send the chunk to a web client if this was a web application
+    def process_audio_chunks(chunks):
+        audio_segments = []
+        for sampling_rate, audio_chunk in chunks:
+            audio_segments.append(audio_chunk)
+            temp_audio_path = os.path.join(tempfile.gettempdir(), f"parler_tts_audio_chunk_{len(audio_segments)}.wav")
+            write_wav(temp_audio_path, sampling_rate, audio_chunk.astype(np.float32))
+            logging.debug(f"Saved chunk to {temp_audio_path}")
+            # Optionally, send this chunk to the client in real-time
+        return audio_segments
+
+    with concurrent.futures.ThreadPoolExecutor() as executor:
+        # Start processing audio chunks in a separate thread
+        future_chunks = executor.submit(process_audio_chunks, generate_chunks(text, description))
+
+        # Continue with other tasks in parallel
+        # (e.g., you can update the chatbot interface, handle other requests, etc.)
+
+        # Wait for audio processing to complete and get the result
+        audio_segments = future_chunks.result()
 
     # Combine all the audio chunks into one audio file
     combined_audio = np.concatenate(audio_segments)
@@ -816,7 +894,6 @@ def generate_audio_parler_tts(text):
 
 
 
-
 def fetch_local_events():
     api_key = os.environ['SERP_API']
     url = f'https://serpapi.com/search.json?engine=google_events&q=Events+in+Birmingham&hl=en&gl=us&api_key={api_key}'