Update app.py

app.py

@@ -12,9 +12,9 @@ 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("sanchit-gandhi/speecht5_tts_vox_nl")
+processor = SpeechT5Processor.from_pretrained("kfahn/speecht5_finetuned_voxpopuli_es")
 
-model = SpeechT5ForTextToSpeech.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl").to(device)
+model = SpeechT5ForTextToSpeech.from_pretrained("kfahn/speecht5_finetuned_voxpopuli_es").to(device)
 vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
 
 embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
@@ -22,7 +22,7 @@ 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": "nl"})
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "es"})
     return outputs["text"]
 
 
@@ -33,9 +33,11 @@ def synthesise(text):
 
 
 def speech_to_speech_translation(audio):
+    target_dtype = np.int16
+    max_range = np.iinfo(target_dtype).max
     translated_text = translate(audio)
     synthesised_speech = synthesise(translated_text)
-    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
+    synthesised_speech = (synthesised_speech.numpy() * max_range).astype(target_dtype)
     return 16000, synthesised_speech