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

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 parse_transcription_with_lm(wav_file):
    speech = convert_file(wav_file)
    inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    with torch.no_grad():
        logits = model(**inputs).logits
    int_result = processor.batch_decode(logits.cpu().numpy())

    transcription =  int_result.text
    return transcription


def convert_file(wav_file):
    filename = wav_file.split('.')[0]
    convert(wav_file, filename + "16k.wav")
    speech, _ = sf.read(filename + "16k.wav")
    return speech

def parse_transcription(wav_file):
    speech = convert_file(wav_file)
    
    
    #speech = read_file(wav_file)
    input_values = processor(speech, sampling_rate=16_000, return_tensors="pt").input_values

    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_transcription, inputs = [input_, chckbox],  outputs=txtbox,
             streaming=True, interactive=True,
             analytics_enabled=False, show_tips=False, enable_queue=True).launch(inline=False);