app.py

import torch
import numpy as np
import soundfile as sf
from transformers import pipeline
from transformers import BarkModel
from transformers import AutoProcessor

device = "cuda:0" if torch.cuda.is_available() else "cpu"

pipe = pipeline(
    "automatic-speech-recognition", model="openai/whisper-large-v2", device=device
)
label = pipeline("audio-classification", model="facebook/mms-lid-126", device=device)
processor = AutoProcessor.from_pretrained("suno/bark")
model = BarkModel.from_pretrained("suno/bark")
model = model.to(device)
synthesised_rate = model.generation_config.sample_rate

def translate(audio_file):
    audio, sampling_rate = sf.read(audio_file)
    outputs = pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe","language":"chinese"})
    language_prediction = label({"array": audio, "sampling_rate": sampling_rate})
    label_outputs = {}
    for pred in language_prediction:
        label_outputs[pred["label"]] = pred["score"]
    return outputs["text"],label_outputs
def synthesise(text_prompt,voice_preset="v2/zh_speaker_1"):
    inputs = processor(text_prompt, voice_preset=voice_preset)
    speech_output = model.generate(**inputs.to(device),pad_token_id=10000)  
    return speech_output
def speech_to_speech_translation(audio,voice_preset="v2/zh_speaker_1"):
    translated_text, label_outputs= translate(audio)
    synthesised_speech = synthesise(translated_text,voice_preset)
    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
    return (synthesised_rate , synthesised_speech.T),translated_text,label_outputs

title = "Multilanguage to Chinese(mandarin) Cascaded STST"
description = """
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in Multilanguage to target speech in Chinese(mandarin). Demo uses OpenAI's [Whisper arge-v2](https://huggingface.co/openai/whisper-large-v2) model for speech translation, and a suno/bark[bark-small](https://huggingface.co/suno/bark) 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")
"""
examples = [
    ["./cs-CZ.mp3", None],
    ["./de-DE.mp3", None],
    ["./es-ES.mp3", None],
    ["./fr-FR.mp3", None],
    ["./it-IT.mp3", None],
    ["./ko-KR.mp3", None],
    ["./nl-NL.mp3", None],
    ["./pl-PL.mp3", None],
    ["./pt-PT.mp3", None],
    ["./ru-RU.mp3", None],
]
import gradio as gr

demo = gr.Blocks()
file_transcribe = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="upload", type="filepath"),
    outputs=[
        gr.Audio(label="Generated Speech", type="numpy"),
        gr.Text(label="Transcription"),
        gr.Label(label="Language prediction"),
    ],
    title=title,
    description=description,
    examples=examples,
)
mic_transcribe = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="microphone", type="filepath"),
    outputs=[
        gr.Audio(label="Generated Speech", type="numpy"),
        gr.Text(label="Transcription"),
        gr.Label(label="Language prediction"),
    ],
    title=title,
    description=description,
)
with demo:
    gr.TabbedInterface(
        [file_transcribe, mic_transcribe],
        ["Transcribe Audio File", "Transcribe Microphone"],
    )

demo.launch(share=True)