bbbipul/vosk-small-streaming-bn-int8

Vosk small streaming Bangla (int8)

An int8-quantized copy of alphacep/vosk-model-small-streaming-bn (version 0.60, revision 501097ae), a streaming Zipformer2 transducer for Bangla speech recognition, for use with sherpa-onnx on phones.

The original model is by Alpha Cephei and is licensed Apache-2.0. This copy is modified: the encoder and joiner were dynamically quantized (MatMul weights to int8) with ONNX Runtime. The decoder and tokens are unchanged.

File	Size
encoder.int8.onnx	26.2 MB
decoder.onnx	2.1 MB
joiner.int8.onnx	0.26 MB
tokens.txt	6 KB

Accuracy

Measured on the FLEURS bn_in dev set (402 utterances, 87 minutes), greedy search, punctuation stripped:

Model	WER	CER
Original (fp32, 90 MB)	26.97	8.84
This copy (int8, 27 MB)	27.06	8.94