app.py

import nltk
import librosa
import torch
import gradio as gr
from pyctcdecode import build_ctcdecoder
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC

nltk.download("punkt")

#Loading the model and the tokenizer
model_name = "facebook/wav2vec2-base-960h"
processor = Wav2Vec2Processor.from_pretrained(model_name)
model = Wav2Vec2ForCTC.from_pretrained(model_name)

def load_data(input_file):  
  #read the file
  speech, sample_rate = librosa.load(input_file)
  #make it 1-D
  if len(speech.shape) > 1: 
      speech = speech[:,0] + speech[:,1]
  #resampling to 16KHz
  if sample_rate !=16000:
    speech = librosa.resample(speech, sample_rate,16000)
  return speech
  
  
def fix_transcription_casing(input_sentence):
  sentences = nltk.sent_tokenize(input_sentence)
  return (' '.join([s.replace(s[0],s[0].capitalize(),1) for s in sentences]))
  
def predict_and_decode(input_file):
  speech = load_data(input_file)
  #tokenize
  input_values = processor(speech, return_tensors="pt", sampling_rate=16000).input_values
  logits = model(input_values).logits.cpu().detach().numpy()[0]
  vocab_list = list(processor.tokenizer.get_vocab().keys())
#   #Take argmax
#   predicted_ids = torch.argmax(logits, dim=-1)
#   #Get the words from predicted word ids
#   transcription = tokenizer.decode(predicted_ids[0])
  decoder = build_ctcdecoder(vocab_list)
  pred = decoder.decode(logits)

  #Output is all upper case
  transcribed_text = fix_transcription_casing(pred.lower())
  return transcribed_text
  
gr.Interface(predict_and_decode,
             inputs = gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Speaker"),
             outputs = gr.outputs.Textbox(label="Output Text"),
             title="ASR using Wav2Vec 2.0 & pyctcdecode",
             description = "Wav2Vec2 in-action",
             layout = "horizontal",
             examples = [["test.wav"]], theme="huggingface").launch()