Indic ASR Quantized

A helper package to use Quantized Indic ASR (Automatic Speech Recognition) for multiple Indic languages.

The original model was developed by AI4Bharat and can be found here

Installation

GPU Installation

pip install uv
uv pip install indic-asr-onnx 

# Specify --extra-index-url if needed

CPU-only Installation (Recommended for limited resources)

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

Quick Start

from indic_asr_onnx import IndicTranscriber

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

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

# Transcribe audio using RNN-T head
text = transcriber.transcribe_rnnt("audio.wav", "hi")  # Hindi
print(text)

Supported Languages

• Assamese (as)
• Bengali (bn)
• Bodo (brx)
• Dogri (doi)
• Gujarati (gu)
• Hindi (hi)
• Kannada (kn)
• Kashmiri (ks)
• Konkani (kok)
• Maithili (mai)
• Malayalam (ml)
• Manipuri (mni)
• Marathi (mr)
• Nepali (ne)
• Odia (or)
• Punjabi (pa)
• Sanskrit (sa)
• Santali (sat)
• Sindhi (sd)
• Tamil (ta)
• Telugu (te)
• Urdu (ur)

Features

• Quantized Models: INT8 quantization for efficient CPU inference
• Multiple Languages: Support for 22 Indic languages
• Two Modes: CTC and RNN-T decoding
• Auto Download: Models download automatically on first use
• ONNX Runtime: Optimized inference with ONNX

Audio Requirements

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