|
import gradio as gr |
|
import numpy as np |
|
import torch |
|
from transliterate import translit |
|
from datasets import load_dataset |
|
|
|
from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline |
|
|
|
|
|
device = "cuda:0" if torch.cuda.is_available() else "cpu" |
|
|
|
|
|
asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device) |
|
|
|
|
|
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts") |
|
|
|
model = SpeechT5ForTextToSpeech.from_pretrained("voxxer/speecht5_finetuned_commonvoice_ru_translit").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) |
|
|
|
|
|
translator = pipeline("translation", model="Helsinki-NLP/opus-mt-en-ru", device=device) |
|
|
|
def translate(audio): |
|
outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"}) |
|
translated_text = translator(outputs["text"]) |
|
return translated_text[0]['translation_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) |
|
return speech.cpu() |
|
|
|
def cleanup_text(inputs): |
|
replacements = [('«', '"'), |
|
('»', '"'), |
|
('‑', '-'), |
|
('–', '-'), |
|
('−', '-'), |
|
('…', '...'), |
|
] |
|
for src, dst in replacements: |
|
inputs = translit(inputs.replace(src, dst).lower(), 'ru', reversed=True) |
|
return inputs |
|
|
|
|
|
def speech_to_speech_translation(audio): |
|
translated_text = translate(audio) |
|
translated_text = cleanup_text(translated_text) |
|
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 Russian. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and finetuned Microsoft's |
|
[SpeechT5 TTS](https://huggingface.co/voxxer/speecht5_finetuned_commonvoice_ru_translit) 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"), |
|
title=title, |
|
description=description, |
|
) |
|
|
|
with demo: |
|
gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"]) |
|
|
|
demo.launch() |
|
|