import soundfile as sf import torch from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor,Wav2Vec2ProcessorWithLM import gradio as gr import sox import subprocess def read_file_and_process(wav_file, processor): filename = wav_file.split('.')[0] filename_16k = filename + "16k.wav" resampler(wav_file, filename_16k) speech, _ = sf.read(filename_16k) inputs = processor(speech, sampling_rate=16_000, return_tensors="pt", padding=True) return inputs def resampler(input_file_path, output_file_path): 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(logits,processor): predicted_ids = torch.argmax(logits, dim=-1) transcription = processor.decode(predicted_ids[0], skip_special_tokens=True) return transcription def parse(wav_file, language): if language == 'Hindi': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-hindi-him-4200") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-hindi-him-4200") elif language == 'Odia': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-odia-orm-100") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-odia-orm-100") elif language == 'Assamese': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-assamese-asm-8") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-assamese-asm-8") elif language == 'Sanskrit': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-sanskrit-sam-60") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-sanskrit-sam-60") elif language == 'Punjabi': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-punjabi-pam-10") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-punjabi-pam-10") elif language == 'Urdu': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-urdu-urm-60") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-urdu-urm-60") elif language == 'Rajasthani': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-rajasthani-raj-45") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-rajasthani-raj-45") elif language == 'Marathi': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-marathi-mrm-100") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-marathi-mrm-100") elif language == 'Malayalam': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-malayalam-mlm-8") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-malayalam-mlm-8") elif language == 'Maithili': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-maithili-maim-50") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-maithili-maim-50") elif language == 'Dogri': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-dogri-doi-55") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-dogri-doi-55") elif language == 'Bhojpuri': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-bhojpuri-bhom-60") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-bhojpuri-bhom-60") elif language == 'Tamil': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-tamil-tam-250") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-tamil-tam-250") elif language == 'Telugu': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-telugu-tem-100") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-telugu-tem-100") elif language == 'Nepali': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-nepali-nem-130") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-nepali-nem-130") elif language == 'Kannada': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-kannada-knm-560") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-kannada-knm-560") elif language == 'Gujarati': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-gujarati-gnm-100") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-gujarati-gnm-100") elif language == 'Bengali': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-bengali-bnm-200") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-bengali-bnm-200") elif language == 'English': processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-indian-english-enm-700") model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-indian-english-enm-700") input_values = read_file_and_process(wav_file, processor) with torch.no_grad(): logits = model(**input_values).logits return parse_transcription(logits, processor) options = ['Hindi','Odia','Assamese','Sanskrit','Punjabi','Urdu','Rajasthani','Marathi','Malayalam','Maithili','Dogri','Bhojpuri','Tamil','Telugu','Nepali','Kannada','Gujarati','Bengali','English'] language = gr.Dropdown(options,label="Select language") input_ = gr.Audio(source="upload", type="filepath") txtbox = gr.Textbox( label="Output from model will appear here:", lines=5 ) gr.Interface(parse, inputs = [input_,language ], outputs=txtbox, streaming=True, interactive=True, analytics_enabled=False, show_tips=False, enable_queue=True).launch(inline=False);