Update app.py

app.py

@@ -14,11 +14,10 @@ device = "cuda:0" if torch.cuda.is_available() else "cpu"
 asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
 
 # load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+processor = SpeechT5Processor.from_pretrained("vadhri/speecht5_finetuned_voxpopuli_nl")
 
 # model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
-model = VitsModel.from_pretrained("facebook/mms-tts-spa").to(device)
-tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-spa")
+model = VitsModel.from_pretrained("vadhri/speecht5_finetuned_voxpopuli_nl").to(device)
 
 vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
 
@@ -27,30 +26,20 @@ speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze
 
 
 def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "es"})
-    print ('Translated text : ', outputs["text"])
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "nl"})
     return outputs["text"]
 
 
 def synthesise(text):
-    inputs = tokenizer(text, return_tensors="pt")
-    
-    with torch.no_grad():
-        speech = model(**inputs)
-
-    print (speech)
-    
-    return speech.waveform
+    inputs = processor(text=text, return_tensors="pt")
+    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
+    return speech.cpu()
 
-def speech_to_speech_translation_fix(audio,voice_preset="v2/zh_speaker_1"):
-    synthesised_rate,synthesised_speech = speech_to_speech_translation(audio,voice_preset)
-    return synthesised_rate,synthesised_speech.T
 
 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 = speech_to_speech_translation_fix(synthesised_speech)
     return 16000, synthesised_speech