app.py

import gradio as gr
import numpy as np
import torch
from datasets import load_dataset
import librosa
from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
from transformers import WhisperForConditionalGeneration, WhisperProcessor
from transformers import VitsModel, VitsTokenizer
    

target_dtype = np.int16
max_range = np.iinfo(target_dtype).max
device = "cuda:0" if torch.cuda.is_available() else "cpu"


model_mms = VitsModel.from_pretrained("facebook/mms-tts-nld")
tokenizer_mms = VitsTokenizer.from_pretrained("facebook/mms-tts-nld")

processor = WhisperProcessor.from_pretrained("openai/whisper-base")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base")
forced_decoder_ids = processor.get_decoder_prompt_ids(language="nl", task="transcribe")
sampling_rate = processor.feature_extractor.sampling_rate


def translate(audio):
    input_features = processor(audio,sampling_rate=sampling_rate,return_tensors="pt").input_features
    predicted_ids = model.generate(input_features,forced_decoder_ids=forced_decoder_ids)
    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
    return transcription

def synthesise(text):
    print("text",text)
    inputs = tokenizer_mms(text[0], return_tensors="pt")
    input_ids = inputs["input_ids"]
    with torch.no_grad():
        outputs = model_mms(input_ids)
    speech = outputs["waveform"]
    return speech
    
def speech_to_speech_translation(audio):
    sampling_rate = 16000
    data_array,samplerate = librosa.load(audio)
    data_16 = librosa.resample(data_array, orig_sr=samplerate, target_sr=sampling_rate)
    translated_text = translate(data_16)
    synthesised_speech = synthesise(translated_text)
    print("max_range",max_range)
    synthesised_speech = (synthesised_speech.numpy() * max_range).astype(np.int16)
    print("sampling_rate",sampling_rate)
    print("synthesised_speech",synthesised_speech)
    return sampling_rate, synthesised_speech.T


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()