language fix + debugging logs

- Changed task language from french to italian
- changed synthesizer to my own
- added some prints for easier debugging

app.py

@@ -7,14 +7,15 @@ from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Proce
 
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
+print("> DEVICE: ", device)
 
 # 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")
+processor = SpeechT5Processor.from_pretrained("jalal-elzein/speecht5_tts_voxpopuli_it_finetuned")
 
-model = SpeechT5ForTextToSpeech.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl").to(device)
+model = SpeechT5ForTextToSpeech.from_pretrained("jalal-elzein/speecht5_tts_voxpopuli_it_finetuned").to(device)
 vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
 
 embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
@@ -22,7 +23,7 @@ speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze
 
 
 def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate", "language": "french"})
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate", "language": "italian"})
     return outputs["text"]
 
 
@@ -34,6 +35,7 @@ def synthesise(text):
 
 def speech_to_speech_translation(audio):
     translated_text = translate(audio)
+    print("> Translated Text: ", translated_text)
     synthesised_speech = synthesise(translated_text)
     synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
     return 16000, synthesised_speech