app.py

import soundfile as sf
import torch
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor,Wav2Vec2ProcessorWithLM
import gradio as gr
import scipy.signal as sps
import sox
import subprocess

def convert(inputfile, outfile):
    sox_tfm = sox.Transformer()
    sox_tfm.set_output_format(
        file_type="wav", channels=1, encoding="signed-integer", rate=16000, bits=16
    )
    #print(this is not done)
    sox_tfm.build(inputfile, outfile)

def read_file(wav):
    sample_rate, signal = wav                                                                                                                        
    signal = signal.mean(-1)                                                                                                                              
    number_of_samples = round(len(signal) * float(16000) / sample_rate)                                                                                   
    resampled_signal = sps.resample(signal, number_of_samples)
    return resampled_signal


def resampler(input_file_path, output_file_path):
    #output_file_path = output_folder_path + input_file_path.split('/')[-1]

    command = (
        f"ffmpeg -hide_banner -loglevel panic -i {input_file_path} -ar 16000 -ac 1 -bits_per_raw_sample 16 -vn "
        f"{output_file_path}"
    )
    subprocess.call(command, shell=True)

def parse_transcription_with_lm(wav_file):
    input_values = read_file_and_process(wav_file)

    # with torch.no_grad():
    #     logits = model(**input_values).logits[0].cpu().numpy()
    # print(logits)
    # int_result = processor_with_LM.decode(logits = logits, output_word_offsets=False,
    #                                      beam_width=128
    #                                      )
    # print(int_result)
    # transcription =  int_result.text.replace('<s>','')


    with torch.no_grad():
        logits = model(**input_values).logits

    result = processor_with_LM.batch_decode(logits.cpu().numpy())
    text = result.text
    transcription = text[0].replace('<s>','')
    return transcription


def read_file_and_process(wav_file):
    filename = wav_file.split('.')[0]
    resampler(wav_file, filename + "16k.wav")
    speech, _ = sf.read(filename + "16k.wav")
    inputs = processor(speech, sampling_rate=16_000, return_tensors="pt", padding=True)
    
    return inputs

def parse(wav_file, applyLM):
    if applyLM:
        return parse_transcription_with_lm(wav_file)
    else:
        return parse_transcription(wav_file)

def parse_transcription(wav_file):
    input_values = read_file_and_process(wav_file)
    with torch.no_grad():
        logits = model(**input_values).logits
    #logits = model(input_values).logits
    predicted_ids = torch.argmax(logits, dim=-1)

    transcription = processor.decode(predicted_ids[0], skip_special_tokens=True)
    return transcription
    
model_id = "Harveenchadha/vakyansh-wav2vec2-hindi-him-4200"
processor = Wav2Vec2Processor.from_pretrained(model_id)
processor_with_LM = Wav2Vec2ProcessorWithLM.from_pretrained(model_id)
model = Wav2Vec2ForCTC.from_pretrained(model_id)
    

    
input_ = gr.Audio(source="microphone", type="filepath") 
#input_ = gr.inputs.Audio(source="microphone", type="numpy") 
txtbox = gr.Textbox(
            label="Output from model will appear here:",
            lines=5
        )

chkbox = gr.Checkbox(label="Apply LM", value=False)

gr.Interface(parse, inputs = [input_, chkbox],  outputs=txtbox,
             streaming=True, interactive=True,
             analytics_enabled=False, show_tips=False, enable_queue=True).launch(inline=False);