app.py

import soundfile as sf
import torch
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer
import gradio as gr
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
    )
    sox_tfm.build(inputfile, outfile)
    
    

model_translate = M2M100ForConditionalGeneration.from_pretrained("facebook/m2m100_418M")
tokenizer_translate = M2M100Tokenizer.from_pretrained("facebook/m2m100_418M")
inlang='hi'
outlang='en'
tokenizer_translate.src_lang = inlang

def translate(text):    
    encoded_hi = tokenizer_translate(text, return_tensors="pt")
    generated_tokens = model_translate.generate(**encoded_hi, forced_bos_token_id=tokenizer_translate.get_lang_id(outlang))
    return tokenizer_translate.batch_decode(generated_tokens, skip_special_tokens=True)[0]
    
    
processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-hindi-him-4200")
model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-hindi-him-4200")

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



output1 = gr.outputs.Textbox(label="Hindi Output from ASR")
output2 = gr.outputs.Textbox(label="English Translated Output")
    
input_ = gr.inputs.Audio(source="microphone", type="file") 
#gr.Interface(parse_transcription, inputs = input_,  outputs="text", 
#             analytics_enabled=False, show_tips=False, enable_queue=True).launch(inline=False);
             
gr.Interface(parse_transcription, inputs = input_,  outputs=[output1, output2], analytics_enabled=False, 
                                                                            show_tips=False, 
                                                                            theme='huggingface',
                                                                            layout='vertical',
                                                                            title="Vakyansh: Speech To text for Indic Languages",
                                                                            description="This is a live demo for Speech to Text Translation. Models used: vakyansh wav2vec2 hindi + m2m100", enable_queue=True).launch( inline=False)