Update app.py

app.py

@@ -1,34 +1,128 @@
 import gradio as gr
 import numpy as np
 import torch
-from datasets import load_dataset
 
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
+from transformers import (
+    VitsModel,
+    VitsTokenizer,
+    pipeline
+)
 
 
 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)
+asr_pipe = pipeline(
+    "automatic-speech-recognition",
+    model="openai/whisper-base",
+    device=device
+)
+
+model = VitsModel.from_pretrained("Matthijs/mms-tts-deu")
+tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-deu")
+
+
+def translate(audio):
+    outputs = asr_pipe(
+        audio,
+        max_new_tokens=256,
+        generate_kwargs={"task": "transcribe", "language": "de"}
+    )
+    return outputs["text"]
+
+
+def synthesise(text):
+    if len(text.strip()) == 0:
+        return (16000, np.zeros(0).astype(np.int16))
+
+    inputs = tokenizer(text, return_tensors="pt")
+    input_ids = inputs["input_ids"]
+
+    with torch.no_grad():
+        outputs = model(input_ids)
+
+    speech = outputs.audio[0]
+    return speech.cpu()
+
+
+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
+
+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:
+![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
+"""
+
+demo = gr.Blocks()
+
+mic_translate = gr.Interface(
+    fn=speech_to_speech_translation,
+    inputs=gr.Audio(source="microphone", type="filepath"),
+    outputs=gr.Audio(label="Generated Speech", type="numpy"),
+    title=title,
+    description=description,
+)
+file_translate = gr.Interface(
+    fn=speech_to_speech_translation,
+    inputs=gr.Audio(source="upload", type="filepath"),
+    outputs=gr.Audio(label="Generated Speech", type="numpy"),
+    examples=[["./example.wav"]],
+    title=title,
+    description=description,
+)
+
+with demo:
+    gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
+
+demo.launch()import gradio as gr
+import numpy as np
+import torch
 
-# load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+from transformers import (
+    VitsModel,
+    VitsTokenizer,
+    pipeline
+)
+
+
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+# load speech translation checkpoint
+asr_pipe = pipeline(
+    "automatic-speech-recognition",
+    model="openai/whisper-base",
+    device=device
+)
 
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
+model = VitsModel.from_pretrained("Matthijs/mms-tts-deu")
+tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-deu")
 
 
 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": "transcribe", "language": "de"}
+    )
     return outputs["text"]
 
 
 def synthesise(text):
-    inputs = processor(text=text, return_tensors="pt")
-    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
+    if len(text.strip()) == 0:
+        return (16000, np.zeros(0).astype(np.int16))
+
+    inputs = tokenizer(text, return_tensors="pt")
+    input_ids = inputs["input_ids"]
+
+    with torch.no_grad():
+        outputs = model(input_ids)
+
+    speech = outputs.audio[0]
     return speech.cpu()
 
 
@@ -38,12 +132,10 @@ def speech_to_speech_translation(audio):
     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:
-
 ![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 +148,6 @@ mic_translate = gr.Interface(
     title=title,
     description=description,
 )
-
 file_translate = gr.Interface(
     fn=speech_to_speech_translation,
     inputs=gr.Audio(source="upload", type="filepath"),
@@ -69,4 +160,4 @@ file_translate = gr.Interface(
 with demo:
     gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
 
-demo.launch()
+demo.launch()