Update app.py

app.py

@@ -2,7 +2,6 @@ import gradio as gr
 import numpy as np
 import torch
 from datasets import load_dataset
-from transformers import VitsModel, VitsTokenizer
 
 from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
 
@@ -11,24 +10,38 @@ 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)
-translator = pipeline("translation", model="Helsinki-NLP/opus-mt-en-ru")
 
+# load text-to-speech checkpoint and speaker embeddings
+tts_checkpoint = "sanchit-gandhi/speecht5_tts_vox_nl"
+processor = SpeechT5Processor.from_pretrained(tts_checkpoint)
 
-model = VitsModel.from_pretrained('facebook/mms-tts-rus').to(device)
-tokenizer = VitsTokenizer.from_pretrained('facebook/mms-tts-rus')
+model = SpeechT5ForTextToSpeech.from_pretrained(tts_checkpoint).to(device)
+vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
 
+embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", 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"})
-    return translator(outputs['text'])[0]['translation_text']
+    # Trick Whisper to translate from any language to Dutch.
+    # Note that using task=translate will translate to English instead.
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "dutch"})
+    return outputs["text"]
 
 
 def synthesise(text):
-    inputs = tokenizer(text=text, return_tensors="pt",max_length=598,
-        truncation=True,)
-    with torch.no_grad():
-        speech = model(**inputs).waveform
+    # Need to specific truncate to max text positions.
+    # Otherwise model.generate_speech will throw errors.
+    inputs = processor(
+        text=text,
+        # max_length=200,
+        max_length=598,
+        truncation=True,
+        # padding=True,
+        return_tensors="pt"
+    )
+    # 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()
 
 
@@ -36,12 +49,13 @@ def speech_to_speech_translation(audio):
     translated_text = translate(audio)
     synthesised_speech = synthesise(translated_text)
     synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
-    return 16000, synthesised_speech[0]
+    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 Russian
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Dutch. Demo uses OpenAI's [Whisper Tiny](https://huggingface.co/openai/whisper-tiny) model for speech translation, and a finetuned
+[SpeechT5 TTS](https://huggingface.co/sanchit-gandhi/speecht5_tts_vox_nl) model 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")
 """