app.py

import gradio as gr
import librosa
from optimum.onnxruntime import ORTModelForSeq2SeqLM
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

# load model and processor
processor = Wav2Vec2Processor.from_pretrained("jonatasgrosman/wav2vec2-large-xlsr-53-english")
model = Wav2Vec2ForCTC.from_pretrained("jonatasgrosman/wav2vec2-large-xlsr-53-english")

tokenizer = AutoTokenizer.from_pretrained("icon-it-tdtu/mt-en-vi-optimum")
model_lm = ORTModelForSeq2SeqLM.from_pretrained("icon-it-tdtu/mt-en-vi-optimum")

def process_audio_file(file):
    data, sr = librosa.load(file)
    if sr != 16000:
        data = librosa.resample(data, sr, 16000)
    inputs = processor(data, sampling_rate=16000, return_tensors="pt", padding=True)
    return inputs


def transcribe(file, state=""):
    inputs = process_audio_file(file)
    with torch.no_grad():
        output_logit = model(inputs.input_values).logits
    pred_ids = torch.argmax(output_logit, dim=-1)
    text = processor.batch_decode(pred_ids)[0].lower()
    print(text)
    text = translate(text)
    state += text + " "
    return state, state


def translate(text):
    batch = tokenizer([text], return_tensors="pt")
    generated_ids = model_lm.generate(**batch)
    translated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    return translated_text


# Set the starting state to an empty string

iface = gr.Interface(
    fn=transcribe,
    title="Streaming interpret English to Vietnamese",
    description="A simple interface to streaming interpret from spoken English to Vietnamese.",
    article="Author: <a href=\"https://huggingface.co./vumichien\">Vu Minh Chien</a>.", 
    inputs=[
        gr.Audio(source="microphone", type="filepath", streaming=True), 
        "state" 
    ],
    outputs=[
        "textbox",
        "state"
    ])
    
iface.launch(enable_queue=True, debug=True)