Update app.py

app.py

@@ -12,6 +12,12 @@ from transformers import (
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
 
+feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-base")
+tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-base", language="french", task="translate")
+
+whisper_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base")
+forced_decoder_ids = tokenizer.get_decoder_prompt_ids(language="french", task="translate")
+
 # load text-to-speech checkpoint and speaker embeddings
 # processor = SpeechT5Processor.from_pretrained("leo-kwan/speecht5_finetuned_voxpopuli_lt")
 
@@ -21,17 +27,14 @@ device = "cuda:0" if torch.cuda.is_available() else "cpu"
 # embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
 # speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
 
+bark_model = BarkModel.from_pretrained("suno/bark-small")
+processor = BarkProcessor.from_pretrained("suno/bark-small")
 
 def translate(audio):
     # load speech translation checkpoint
-    feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-base")
-    tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-base", language="french", task="translate")
-    
-    model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base")
-    forced_decoder_ids = tokenizer.get_decoder_prompt_ids(language="french", task="translate")
     asr_pipe = pipeline(
         "automatic-speech-recognition", 
-        model=model, 
+        model=whisper_model, 
         feature_extractor=feature_extractor,
         tokenizer=tokenizer,
         device=device
@@ -41,10 +44,9 @@ def translate(audio):
 
 
 def synthesise(text):
-    model = BarkModel.from_pretrained("suno/bark-small")
-    processor = BarkProcessor.from_pretrained("suno/bark-small")
+    
     inputs = processor(text, voice_preset="v2/fr_speaker_1")
-    speech = model.generate(**inputs).cpu()
+    speech = bark_model.generate(**inputs).cpu()
     
     # inputs = processor(text=text, return_tensors="pt")
     # speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
@@ -54,7 +56,7 @@ def synthesise(text):
 def speech_to_speech_translation(audio):
     translated_text = translate(audio)
     synthesised_speech = synthesise(translated_text)
-    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
+    # synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
     return 16000, synthesised_speech