Use lt fine-tuned model for TTS

app.py

@@ -11,10 +11,13 @@ 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("microsoft/speecht5_tts")
+# load text-to-speech checkpoint and speaker embeddings 
+#processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+processor = SpeechT5Processor.from_pretrained("ihanif/speecht5_finetuned_voxpopuli_lt")
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
+#model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
+#vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+model = SpeechT5ForTextToSpeech.from_pretrained("ihanif/speecht5_finetuned_voxpopuli_lt").to(device)
 vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
 
 embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
@@ -22,7 +25,7 @@ speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze
 
 
 def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate", "language": "lt"})
     return outputs["text"]