app.py

import torch
from transformers import pipeline

device="cpu"
pipe = pipeline(
    "automatic-speech-recognition", model="openai/whisper-large-v2", device=device
)

def translate(audio):
    outputs = pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe","language":"chinese"})
    return outputs["text"]

from transformers import BarkModel
from transformers import AutoProcessor
model = BarkModel.from_pretrained("suno/bark-small")
processor = AutoProcessor.from_pretrained("suno/bark")

model = model.to(device)
synthesised_rate = model.generation_config.sample_rate


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)  
    #print(speech_output[0].cpu().numpy())
    return speech_output


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

import numpy as np
def speech_to_speech_translation(audio,voice_preset="v2/zh_speaker_1"):
    translated_text = translate(audio)
    #print(translated_text)
    synthesised_speech = synthesise(translated_text,voice_preset)
    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
    #print(synthesised_speech)  
    return synthesised_rate , synthesised_speech 
def speech_to_speech_translation_fix(audio,voice_preset="v2/zh_speaker_1"):
    synthesised_rate,synthesised_speech = speech_to_speech_translation(audio,voice_preset)
    return synthesised_rate,synthesised_speech.T

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")
"""

demo = gr.Blocks()

mic_translate = gr.Interface(
    fn=speech_to_speech_translation_fix,
    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_fix,
    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()