Update app.py

app.py

@@ -11,37 +11,25 @@ target_dtype = np.int16
 max_range = np.iinfo(target_dtype).max
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
-# load speech translation checkpoint
-
-asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
-
-# load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl")
-
-model = SpeechT5ForTextToSpeech.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl").to(device)
-
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
-
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
 
 model_mms = VitsModel.from_pretrained("facebook/mms-tts-nld")
 tokenizer_mms = VitsTokenizer.from_pretrained("facebook/mms-tts-nld")
 
-
-def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"language": "nl","task": "transcribe"})
-    return outputs["text"]
+processor = WhisperProcessor.from_pretrained("openai/whisper-base")
+model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base").to(device)
+forced_decoder_ids = processor.get_decoder_prompt_ids(language="nl", task="transcribe")
+sampling_rate = processor.feature_extractor.sampling_rate
 
 
-def synthesise_speechT5(text):
-    inputs = processor(text=text, padding='max_length', truncation=True,max_length=600,return_tensors="pt")
-    print(inputs)
-    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device),vocoder=vocoder)
-    return speech.cpu()
+def translate(audio):
+    input_features = processor(audio,sampling_rate=sampling_rate,return_tensors="pt").input_features
+    predicted_ids = model.generate(input_features.to(device),forced_decoder_ids=forced_decoder_ids)
+    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
+    return transcription
 
 def synthesise(text):
-    inputs = tokenizer_mms(text, return_tensors="pt")
+    print(text)
+    inputs = tokenizer_mms(text[0], return_tensors="pt")
     input_ids = inputs["input_ids"]
     with torch.no_grad():
         outputs = model_mms(input_ids)