vakgyata is an open-source language identification model specifically designed to classify Indian languages from raw speech audio. It is built upon the pretrained Harveenchadha/wav2vec2-pretrained-clsril-23-10k with additional Layer Normalization integrated to improve stability and performance for audio classification tasks.

Variants and Model Sizes

Variant	Parameters	Accuracy
`vakgyata-base`	95M	95.88%
`vakgyata-small`	52M	95.06%
`vakgyata-mini`	38M	95.06%
`vakgyata-tiny`	24M	93.63%

Supported Languages

Language	Code
English (India)	en-IN
Hindi	hi-IN
Odia	or-IN
Bengali	bn-IN
Tamil	ta-IN
Telugu	te-IN
Kannada	kn-IN
Malayalam	ml-IN
Marathi	mr-IN
Gujarati	gu-IN
Punjabi	pa-IN
Assamese	as-IN

Specifications

Supported Sampling Rate: 16000 Hz
Recommended Audio Format: 16kHz, 16bit PCM (Mono)

Installation

pip install transformers torchaudio

Usage

python

from transformers import Wav2Vec2ForSequenceClassification, AutoFeatureExtractor
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"

model_id = "onecxi/vakgyata-small"

processor = AutoFeatureExtractor.from_pretrained(model_id)
model = Wav2Vec2ForSequenceClassification.from_pretrained(model_id).to(device)

Inference Example

python

import torchaudio

# Load the audio (ensure it's 16kHz mono)
audio, sr = torchaudio.load("path/to/audio.wav")

# Preprocess
inputs = processor(audio.squeeze(), sampling_rate=sr, return_tensors="pt").to(device)

# Inference
with torch.no_grad():
    logits = model(**inputs).logits

# Softmax to get probabilities
probs = logits.softmax(dim=-1).cpu().numpy()

# Predicted language
language = model.config.id2label.get(probs.argmax())
print("Predicted Language:", language)

Citation

If you use this model in your research or application, please consider citing the model and its base source:

@misc{vakgyata2024,
  title={vakgyata: Language Identification for Indian Speech},
  author={OneCXI},
  year={2024},
  url={https://huggingface.co/onecxi/vakgyata-small}
}