app.py

import gradio as gr

import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import pandas as pd
from sklearn.model_selection import train_test_split



processor = Wav2Vec2Processor.from_pretrained("gagan3012/wav2vec2-xlsr-nepali")
model = Wav2Vec2ForCTC.from_pretrained("gagan3012/wav2vec2-xlsr-nepali")

from torchaudio.transforms import Resample
import numpy as np





def transcribe_audio(audio_file):
  input_arr, sampling_rate =torchaudio.load(audio_file)
  resampler = Resample(orig_freq=sampling_rate, new_freq=16000)
  input_arr = resampler(input_arr).squeeze().numpy()
  sampling_rate = 16000
  inputs = processor(input_arr, sampling_rate=16_000, return_tensors="pt", padding=True)

  with torch.no_grad():
    logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits

  predicted_ids = torch.argmax(logits, dim=-1)

  predicted_words=  processor.batch_decode(predicted_ids)

  return predicted_words[0]

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

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

iface.launch(inline=False)