Update app.py

app.py

@@ -3,9 +3,7 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
-# from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, 
-from transformers import pipeline
-from transformers import VitsModel, VitsTokenizer
+from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
 
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
@@ -14,28 +12,37 @@ 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")
-# model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
-# vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+tts_checkpoint = "sanchit-gandhi/speecht5_tts_vox_nl"
+processor = SpeechT5Processor.from_pretrained(tts_checkpoint)
 
-# load text-to-speach checkpoint for german language
-model = VitsModel.from_pretrained("Matthijs/mms-tts-deu")
-tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-deu")
+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", "language": "de"})
+    # 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")
-    with torch.no_grad():
-      speech = model(inputs["input_ids"].to(device))
-    return speech.audio[0]
+    # 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()
 
 
 def speech_to_speech_translation(audio):
@@ -47,8 +54,8 @@ def speech_to_speech_translation(audio):
 
 title = "Cascaded STST"
 description = """
-Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in German. 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 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")
 """
 
@@ -56,7 +63,7 @@ demo = gr.Blocks()
 
 mic_translate = gr.Interface(
     fn=speech_to_speech_translation,
-    inputs=gr.Audio(sources="microphone", type="filepath"),
+    inputs=gr.Audio(label="microphone", type="filepath"),
     outputs=gr.Audio(label="Generated Speech", type="numpy"),
     title=title,
     description=description,
@@ -64,7 +71,7 @@ mic_translate = gr.Interface(
 
 file_translate = gr.Interface(
     fn=speech_to_speech_translation,
-    inputs=gr.Audio(sources="upload", type="filepath"),
+    inputs=gr.Audio(label="upload", type="filepath"),
     outputs=gr.Audio(label="Generated Speech", type="numpy"),
     examples=[["./example.wav"]],
     title=title,