ai4bharat / MahaDhwani_pretrained_conformer

MahaDhwani Pretrained Conformer

It is a self-supervised pre-trained conformer encoder model trained on MahaDhwani dataset.

Language

Contains training data from 22 scheduled languages of India.

Input

This model accepts 16000 KHz Mono-channel Audio (wav files) as input.

Output

This model provides conformer encoder embeddings as the output for a given audio sample.

Model Architecture

This model is a conformer-Large model, consisting of 120M parameters, as the encoder. The model has 17 conformer blocks with 512 as the model dimension.

AI4Bharat NeMo:

To load, train, fine-tune or play with the model you will need to install AI4Bharat NeMo . We recommend you install it using the command shown below

git clone https://github.com/AI4Bharat/NeMo.git && cd NeMo && git checkout nemo-v2 && bash reinstall.sh

Usage

Download and load the model from Huggingface.

import pydub
import numpy as np
import torch
import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.from_pretrained("ai4bharat/MahaDhwani_pretrained_conformer")

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.freeze() # inference mode
model = model.to(device) # transfer model to device

Get an audio file ready by running the command shown below in your terminal. This will convert the audio to 16000 Hz and monochannel.

ffmpeg -i sample_audio.wav -ac 1 -ar 16000 sample_audio_infer_ready.wav

Inference

wavpath = 'sample.wav'
wav = pydub.AudioSegment.from_file(wavpath).set_frame_rate(16000).set_channels(1)
sarray = wav.get_array_of_samples()
fp_arr = np.array(sarray).T.astype(np.float64)
fp_arr = fp_arr.reshape((1,-1))
feature = torch.from_numpy(fp_arr).float().to(device='cuda')
length=torch.tensor([fp_arr.shape[1]]).to(device='cuda')

spectrograms, spec_masks, encoded, encoded_len = model(input_signal=feature,input_signal_length=length)