Spaces:
Build error
Build error
Update app.py
Browse files
app.py
CHANGED
@@ -755,18 +755,87 @@ repo_id = "parler-tts/parler-tts-mini-v1"
|
|
755 |
|
756 |
|
757 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
758 |
def generate_audio_parler_tts(text):
|
759 |
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."
|
760 |
-
|
|
|
761 |
|
762 |
# Initialize the tokenizer and model
|
763 |
parler_tokenizer = AutoTokenizer.from_pretrained(repo_id)
|
764 |
parler_model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id).to(device)
|
|
|
765 |
sampling_rate = parler_model.audio_encoder.config.sampling_rate
|
766 |
frame_rate = parler_model.audio_encoder.config.frame_rate
|
|
|
|
|
767 |
|
768 |
-
def
|
769 |
-
play_steps = int(frame_rate * play_steps_in_s)
|
770 |
streamer = ParlerTTSStreamer(parler_model, device=device, play_steps=play_steps)
|
771 |
|
772 |
inputs = parler_tokenizer(description, return_tensors="pt").to(device)
|
@@ -779,7 +848,7 @@ def generate_audio_parler_tts(text):
|
|
779 |
prompt_attention_mask=prompt.attention_mask,
|
780 |
streamer=streamer,
|
781 |
do_sample=True,
|
782 |
-
temperature=
|
783 |
min_new_tokens=10,
|
784 |
)
|
785 |
|
@@ -789,18 +858,27 @@ def generate_audio_parler_tts(text):
|
|
789 |
for new_audio in streamer:
|
790 |
if new_audio.shape[0] == 0:
|
791 |
break
|
792 |
-
# Save or process each audio chunk as it is generated
|
793 |
yield sampling_rate, new_audio
|
794 |
|
795 |
-
|
796 |
-
|
797 |
-
|
798 |
-
|
799 |
-
|
800 |
-
|
801 |
-
|
802 |
-
|
803 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
804 |
|
805 |
# Combine all the audio chunks into one audio file
|
806 |
combined_audio = np.concatenate(audio_segments)
|
@@ -816,7 +894,6 @@ def generate_audio_parler_tts(text):
|
|
816 |
|
817 |
|
818 |
|
819 |
-
|
820 |
def fetch_local_events():
|
821 |
api_key = os.environ['SERP_API']
|
822 |
url = f'https://serpapi.com/search.json?engine=google_events&q=Events+in+Birmingham&hl=en&gl=us&api_key={api_key}'
|
|
|
755 |
|
756 |
|
757 |
|
758 |
+
# def generate_audio_parler_tts(text):
|
759 |
+
# 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."
|
760 |
+
# chunk_size_in_s = 0.5
|
761 |
+
|
762 |
+
# # Initialize the tokenizer and model
|
763 |
+
# parler_tokenizer = AutoTokenizer.from_pretrained(repo_id)
|
764 |
+
# parler_model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id).to(device)
|
765 |
+
# sampling_rate = parler_model.audio_encoder.config.sampling_rate
|
766 |
+
# frame_rate = parler_model.audio_encoder.config.frame_rate
|
767 |
+
|
768 |
+
# def generate(text, description, play_steps_in_s=0.5):
|
769 |
+
# play_steps = int(frame_rate * play_steps_in_s)
|
770 |
+
# streamer = ParlerTTSStreamer(parler_model, device=device, play_steps=play_steps)
|
771 |
+
|
772 |
+
# inputs = parler_tokenizer(description, return_tensors="pt").to(device)
|
773 |
+
# prompt = parler_tokenizer(text, return_tensors="pt").to(device)
|
774 |
+
|
775 |
+
# generation_kwargs = dict(
|
776 |
+
# input_ids=inputs.input_ids,
|
777 |
+
# prompt_input_ids=prompt.input_ids,
|
778 |
+
# attention_mask=inputs.attention_mask,
|
779 |
+
# prompt_attention_mask=prompt.attention_mask,
|
780 |
+
# streamer=streamer,
|
781 |
+
# do_sample=True,
|
782 |
+
# temperature=1.0,
|
783 |
+
# min_new_tokens=10,
|
784 |
+
# )
|
785 |
+
|
786 |
+
# thread = Thread(target=parler_model.generate, kwargs=generation_kwargs)
|
787 |
+
# thread.start()
|
788 |
+
|
789 |
+
# for new_audio in streamer:
|
790 |
+
# if new_audio.shape[0] == 0:
|
791 |
+
# break
|
792 |
+
# # Save or process each audio chunk as it is generated
|
793 |
+
# yield sampling_rate, new_audio
|
794 |
+
|
795 |
+
# audio_segments = []
|
796 |
+
# for (sampling_rate, audio_chunk) in generate(text, description, chunk_size_in_s):
|
797 |
+
# audio_segments.append(audio_chunk)
|
798 |
+
# # Here, you can save the chunk to a file or send it to a frontend
|
799 |
+
# # For example, you could write the chunk to a file immediately:
|
800 |
+
# temp_audio_path = os.path.join(tempfile.gettempdir(), f"parler_tts_audio_chunk_{len(audio_segments)}.wav")
|
801 |
+
# write_wav(temp_audio_path, sampling_rate, audio_chunk.astype(np.float32))
|
802 |
+
# logging.debug(f"Saved chunk to {temp_audio_path}")
|
803 |
+
# # You could also send the chunk to a web client if this was a web application
|
804 |
+
|
805 |
+
# # Combine all the audio chunks into one audio file
|
806 |
+
# combined_audio = np.concatenate(audio_segments)
|
807 |
+
# combined_audio_path = os.path.join(tempfile.gettempdir(), "parler_tts_combined_audio_stream.wav")
|
808 |
+
|
809 |
+
# write_wav(combined_audio_path, sampling_rate, combined_audio.astype(np.float32))
|
810 |
+
|
811 |
+
# logging.debug(f"Combined audio saved to {combined_audio_path}")
|
812 |
+
# return combined_audio_path
|
813 |
+
|
814 |
+
import concurrent.futures
|
815 |
+
import tempfile
|
816 |
+
import os
|
817 |
+
import numpy as np
|
818 |
+
from threading import Thread
|
819 |
+
from transformers import AutoTokenizer
|
820 |
+
from parler_tts import ParlerTTSForConditionalGeneration, ParlerTTSStreamer
|
821 |
+
from scipy.io.wavfile import write as write_wav
|
822 |
+
import logging
|
823 |
+
|
824 |
def generate_audio_parler_tts(text):
|
825 |
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."
|
826 |
+
|
827 |
+
chunk_size_in_s = 0.3 # Smaller chunk size for lower latency
|
828 |
|
829 |
# Initialize the tokenizer and model
|
830 |
parler_tokenizer = AutoTokenizer.from_pretrained(repo_id)
|
831 |
parler_model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id).to(device)
|
832 |
+
|
833 |
sampling_rate = parler_model.audio_encoder.config.sampling_rate
|
834 |
frame_rate = parler_model.audio_encoder.config.frame_rate
|
835 |
+
|
836 |
+
play_steps = int(frame_rate * chunk_size_in_s)
|
837 |
|
838 |
+
def generate_chunks(text, description):
|
|
|
839 |
streamer = ParlerTTSStreamer(parler_model, device=device, play_steps=play_steps)
|
840 |
|
841 |
inputs = parler_tokenizer(description, return_tensors="pt").to(device)
|
|
|
848 |
prompt_attention_mask=prompt.attention_mask,
|
849 |
streamer=streamer,
|
850 |
do_sample=True,
|
851 |
+
temperature=0.7, # Lower temperature for faster generation
|
852 |
min_new_tokens=10,
|
853 |
)
|
854 |
|
|
|
858 |
for new_audio in streamer:
|
859 |
if new_audio.shape[0] == 0:
|
860 |
break
|
|
|
861 |
yield sampling_rate, new_audio
|
862 |
|
863 |
+
def process_audio_chunks(chunks):
|
864 |
+
audio_segments = []
|
865 |
+
for sampling_rate, audio_chunk in chunks:
|
866 |
+
audio_segments.append(audio_chunk)
|
867 |
+
temp_audio_path = os.path.join(tempfile.gettempdir(), f"parler_tts_audio_chunk_{len(audio_segments)}.wav")
|
868 |
+
write_wav(temp_audio_path, sampling_rate, audio_chunk.astype(np.float32))
|
869 |
+
logging.debug(f"Saved chunk to {temp_audio_path}")
|
870 |
+
# Optionally, send this chunk to the client in real-time
|
871 |
+
return audio_segments
|
872 |
+
|
873 |
+
with concurrent.futures.ThreadPoolExecutor() as executor:
|
874 |
+
# Start processing audio chunks in a separate thread
|
875 |
+
future_chunks = executor.submit(process_audio_chunks, generate_chunks(text, description))
|
876 |
+
|
877 |
+
# Continue with other tasks in parallel
|
878 |
+
# (e.g., you can update the chatbot interface, handle other requests, etc.)
|
879 |
+
|
880 |
+
# Wait for audio processing to complete and get the result
|
881 |
+
audio_segments = future_chunks.result()
|
882 |
|
883 |
# Combine all the audio chunks into one audio file
|
884 |
combined_audio = np.concatenate(audio_segments)
|
|
|
894 |
|
895 |
|
896 |
|
|
|
897 |
def fetch_local_events():
|
898 |
api_key = os.environ['SERP_API']
|
899 |
url = f'https://serpapi.com/search.json?engine=google_events&q=Events+in+Birmingham&hl=en&gl=us&api_key={api_key}'
|