indic-asr-onnx 0.1.0

Quantized IndicConformer ASR for multiple Indian languages

Indic ASR

A quantized automatic speech recognition (ASR) system for multiple Indic languages using the IndicConformer architecture.

Installation

CPU-only Installation (Recommended for limited resources)

pip install --extra-index-url https://download.pytorch.org/whl/cpu indic-asr

GPU Installation

pip install indic-asr

For uv users:

uv pip install --extra-index-url https://download.pytorch.org/whl/cpu indic-asr

Quick Start

from indic_asr import IndicConformerTranscriber

# Initialize (downloads model automatically)
transcriber = IndicConformerTranscriber()

# Transcribe audio
text = transcriber.transcribe_ctc("audio.wav", "hi")  # Hindi
print(text)

Supported Languages

• Hindi (hi)
• Bengali (bn)
• Telugu (te)
• Marathi (mr)
• Tamil (ta)
• Gujarati (gu)
• Kannada (kn)
• Malayalam (ml)
• Odia (or)
• Punjabi (pa)
• Assamese (as)

Features

• Quantized Models: INT8 quantization for efficient CPU inference
• Multiple Languages: Support for 11 Indic languages
• Two Modes: CTC and RNN-T decoding
• Auto Download: Models download automatically on first use
• ONNX Runtime: Optimized inference with ONNX

Usage

CTC Mode (Faster)

text = transcriber.transcribe_ctc("audio.wav", "hi")

RNN-T Mode (More Accurate)

text = transcriber.transcribe_rnnt("audio.wav", "hi")

Audio Requirements

• Format: WAV, MP3, FLAC, etc.
• Sample Rate: Auto-resampled to 16kHz
• Channels: Mono (auto-converted)

Performance

• CPU Inference: ~50-100x real-time on modern CPUs
• Memory: ~200-500MB per inference
• Model Size: ~500MB (downloaded on first use)

License

MIT License

Citation

If you use this in your research, please cite:

@misc{indic-asr-2025,
  title={Indic ASR: Quantized Conformer for Indic Languages},
  author={Verma, Atharva},
  year={2025}
}