app.py

import gradio as gr
import numpy as np
import torch
from transformers import AutoProcessor, pipeline, BarkModel

ASR_MODEL_NAME = "bofenghuang/whisper-large-v2-cv11-german"
TTS_MODEL_NAME = "suno/bark-small"
BATCH_SIZE = 8
voices = {
"male" : "v2/de_speaker_0",
"female" : "v2/de_speaker_3"
}

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

# load speech translation checkpoint
asr_pipe = pipeline("automatic-speech-recognition", model=ASR_MODEL_NAME, chunk_length_s=10,device=device)
# set the forced ids
asr_pipe.model.config.forced_decoder_ids = asr_pipe.tokenizer.get_decoder_prompt_ids(language='de', task="translate")

# load text-to-speech checkpoint
processor = AutoProcessor.from_pretrained("suno/bark-small")
model = BarkModel.from_pretrained("suno/bark-small").to(device)
sampling_rate = model.generation_config.sample_rate

def translate(audio):
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
    return outputs["text"]
    
def synthesise(text, voice_preset):
    inputs = processor(text=text, return_tensors="pt",voice_preset=voice_preset)
    speech = model.generate(**inputs.to(device))
    return speech[0]
    
def speech_to_speech_translation(audio, voice):
    voice_preset = None
    translated_text = translate(audio)
    print(translated_text)
    if voice == "Female":
     voice_preset = voices["female"]
    else:
     voice_preset = voices["male"]
    synthesised_speech = synthesise(translated_text, voice_preset)
    synthesised_speech = (synthesised_speech.cpu().numpy() * 32767).astype(np.int16)
    return sampling_rate, synthesised_speech
    
title = "Cascaded STST - Any language to German speech"
description = """
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in German. Demo uses fine-tuned version of openai/whisper-large-v2 model (https://huggingface.co/bofenghuang/whisper-large-v2-cv11-german) for speech translation, and Suno's
[Bark-large](https://huggingface.co/suno/bark-small) 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"),
    gr.inputs.Radio(["Male", "Female"], label="Voice", default="Male")],
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    title=title,
    description=description,
    allow_flagging="never"
)

file_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=[gr.Audio(source="upload", type="filepath"),
    gr.inputs.Radio(["Male", "Female"], label="Voice", default="Male")],
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    title=title,
    description=description,
    allow_flagging="never"
)

with demo:
    gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])

demo.queue(concurrency_count=2,max_size=10)
demo.launch()