Updated app.py

Added the modifications to translate and synthesize method to convert into Spanish and then synthesize using the SPeechT5 model on Dutch language

app.py

@@ -12,7 +12,8 @@ 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("microsoft/speecht5_tts")
+#processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+processor = SpeechT5Processor.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl")
 
 model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
 vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
@@ -22,11 +23,12 @@ 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 = pipe(audio, max_new_tokens=256, generate_kwargs={"task":"translate"})
+    outputs = pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "es"})
     return outputs["text"]
 
 
-def synthesise(text):
+def synthesize(text):
     inputs = processor(text=text, return_tensors="pt")
     speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
     return speech.cpu()
@@ -34,15 +36,15 @@ def synthesise(text):
 
 def speech_to_speech_translation(audio):
     translated_text = translate(audio)
-    synthesised_speech = synthesise(translated_text)
+    synthesised_speech = synthesize(translated_text)
     synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
     return 16000, synthesised_speech
 
 
 title = "Cascaded STST"
 description = """
-Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in English. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
-[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in English. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Fine tuned version of Microsoft's
+[SpeechT5 TTS](https://huggingface.co/sanchit-gandhi/speecht5_tts_vox_nl) model by [Sanchit Gandhi](https://huggingface.co/sanchit-gandhi) for text-to-speech:
 
 ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
 """