Rstc989/Vaanilipi

IndicConformer 600M multilingual, int8

A smaller copy of IndicConformer 600M multilingual by AI4Bharat (IIT Madras): one speech recognition model for the 22 scheduled languages of India. It is used by VaaniLipi, a Windows app for private, offline voice typing.

What is changed

Only the encoder, quantized to int8 in its matrix multiplications (ONNX Runtime dynamic quantization, per channel): 882 MB instead of 2.5 GB and about twice as fast on a CPU. Everything else is AI4Bharat's, unchanged.

File	What it is
encoder.onnx	the Conformer encoder, int8 in its MatMul layers
ctc_decoder.onnx	the CTC head over the shared vocabulary (unchanged)
vocab.json, language_masks.json	each language's tokens and its slice of the vocabulary (unchanged)
mel_fb.bin	the 257 × 80 mel filterbank of the original preprocessor (float32, little endian)
LICENSE	AI4Bharat's MIT licence

Credit and licence

IndicConformer is made by AI4Bharat and released under the MIT licence; this copy keeps that licence (see LICENSE). Please cite and credit the original model. This copy is not made or endorsed by AI4Bharat.