app.py

# -*- coding: utf-8 -*-
"""Untitled29.ipynb

Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/drive/1Lv3LjRH9bHwMhKsWvFcELMzKqmXd9UIb
"""

!pip install -q transformers
!pip install -q gradio

import nltk
import librosa
import torch
import soundfile as sf
import gradio as gr
from transformers import Wav2Vec2Tokenizer, Wav2Vec2ForCTC
nltk.download("punkt")

input_file = "/content/drive/MyDrive/AAAAUDIO/My Audio.wav"

tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-base-960h")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")

def load_data(input_file):
  
  """ Function for resampling to ensure that the speech input is sampled at 16KHz.
  """
  #read the file
  speech, sample_rate = sf.read(input_file)

  #make it 1-D
  if len(speech.shape) > 1: 
      speech = speech[:,0] + speech[:,1]
  
  #Resampling at 16KHz since wav2vec2-base-960h is pretrained and fine-tuned on speech audio sampled at 16 KHz.
  if sample_rate !=16000:
    speech = librosa.resample(speech, sample_rate,16000)
  return speech

def asr_transcript(input_file):
  speech = load_data(input_file)

  #Tokenize
  input_values = tokenizer(speech, return_tensors="pt").input_values

  #Take logits
  logits = model(input_values).logits

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

  #Get the words from predicted word ids
  transcription = tokenizer.decode(predicted_ids[0])

  #Output is all upper case
  transcription = correct_casing(transcription.lower())

  return transcription

gr.Interface(asr_transcript,
             inputs = gr.inputs.Audio(label = "Input Audio", type= "file"),
             outputs = gr.outputs.Textbox(label="Output Text"),
             title="Real-time ASR using Wav2Vec 2.0",
             description = "asdfghnjmk",
             examples = [["/content/drive/MyDrive/AAAAUDIO/My Audio.wav"]]).launch()