Update app.py

app.py

@@ -7,22 +7,32 @@ from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Proce
 
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
+language = "it" # italian 
 
 # load speech translation checkpoint
 asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
 
+# model_tts_name = "microsoft/speecht5_tts"
+# model_vocoder = "microsoft/speecht5_hifigan"
+# model_embedding = "Matthijs/cmu-arctic-xvectors"
+model_tts_name = "menevsem/speecht5_finetuned_voxpopuli_it"
+model_vocoder = "microsoft/speecht5_hifigan"
+model_embedding = "Matthijs/cmu-arctic-xvectors"
+
 # load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+processor = SpeechT5Processor.from_pretrained(model_tts_name)
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+model = SpeechT5ForTextToSpeech.from_pretrained(model_tts_name).to(device)
+vocoder = SpeechT5HifiGan.from_pretrained(model_vocoder).to(device)
 
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
+embeddings_dataset = load_dataset(model_embedding, split="validation")
 speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
 
 
 def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
+    # generate_kwargs = {"task":"translate"}
+    generate_kwargs = {"task": "transcribe", "language": language}
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs=generate_kwargs)
     return outputs["text"]