app.py

import gradio as gr
import librosa
import logging
import numpy as np
import torch

from transformers import VitsModel, VitsTokenizer, pipeline
from transformers import WhisperForConditionalGeneration, WhisperProcessor


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

target_language = "fr"

# load speech translation checkpoint
asr_pipe = pipeline("automatic-speech-recognition", model="bofenghuang/whisper-small-cv11-french", device=device)
# whisper_model_name = "openai/whisper-small"
# whisper_processor = WhisperProcessor.from_pretrained(whisper_model_name)
# whisper_model = WhisperForConditionalGeneration.from_pretrained(whisper_model_name)
# decoder_ids = whisper_processor.get_decoder_prompt_ids(language=target_language, task="transcribe")

# load text-to-speech checkpoint
model = VitsModel.from_pretrained("facebook/mms-tts-fra")
tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-fra")


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

# def translate(audio):
#     if isinstance(audio, str):
#         # Account for recorded audio
#         audio = {
#             "path": audio,
#             "sampling_rate": 16_000,
#             "array": librosa.load(audio, sr=16_000)[0]
#         }
#     elif audio["sampling_rate"] != 16_000:
#         audio["array"] = librosa.resample(audio["array"], audio["sampling_rate"], 16_000)
#     input_features = whisper_processor(audio["array"], sampling_rate=16000, return_tensors="pt").input_features
#     predicted_ids = whisper_model.generate(input_features, forced_decoder_ids=decoder_ids)
#     translated_text = whisper_processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
#     return translated_text


def synthesise(text):
    inputs = tokenizer(text, return_tensors="pt")
    with torch.no_grad():
        outputs = model(inputs["input_ids"])
    speech = outputs["waveform"][0]
    logging.info(speech)
    return speech.cpu()


def speech_to_speech_translation(audio):
    translated_text = translate(audio)
    logging.info(f"Translated 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 French. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
[SpeechT5 TTS](https://huggingface.co/preetam8/speecht5_finetuned_voxpopuli_fr) model for text-to-speech finetuned for french:

![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"])

logging.getLogger().setLevel(logging.INFO)
demo.launch()