app.py

import gradio as gr
import librosa
import torch
import torchaudio
from datasets import load_dataset
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
import pandas as pd
from sklearn.model_selection import train_test_split

from noisereduce.torchgate import TorchGate as TG
import re
from pydub import AudioSegment



from torchaudio.transforms import Resample
import numpy as np



def transcribe_audio(audio_file):
    audio = AudioSegment.from_wav(audio_file)

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

    input_arr, sampling_rate =librosa.load(audio_file)
    # Create TorchGating instance
    tg = TG(sr=sampling_rate, nonstationary=True).to(device)
    try:
      input_arr = tg(input_arr)
    except:
      input_arr = input_arr
    if sampling_rate != 16000:
      input_arr = librosa.resample(input_arr, orig_sr=sampling_rate, target_sr=16000)

    
    MODEL_NAME = "rikeshsilwalekg/whisper-small-wer35-ekg"

    torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

    model = AutoModelForSpeechSeq2Seq.from_pretrained(
        MODEL_NAME, torch_dtype=torch_dtype, use_safetensors=True
    )
    model.to(device)

    processor = AutoProcessor.from_pretrained(MODEL_NAME)

    pipe = pipeline(
        "automatic-speech-recognition",
        model=model,
        tokenizer=processor.tokenizer,
        feature_extractor=processor.feature_extractor,
        max_new_tokens=128,
        chunk_length_s=30,
        batch_size=16,
        return_timestamps=False,
        torch_dtype=torch_dtype,
        device=device,
    )

    # return_timestamps=True for sentence level timestaps
    # for word level timestamps return_timestamps="word"
    prediction = pipe(input_arr)

    
    prediction = prediction['text']

audio_input = gr.inputs.Audio(source="upload", type="filepath")

iface = gr.Interface(fn=transcribe_audio, inputs=audio_input,
                         outputs=["textbox"], title="Nepali Speech To Text",
                         description="Upload an audio file and hit the 'Submit'\
                             button")

iface.launch(inline=False)