krut42/voice-zipformer-tg-int8

Vosk Tajik Zipformer (0.61), sherpa-onnx int8

The Tajik speech recognition model that the «Слышно» Android app downloads for on-device transcription. This repository is the fallback source for the app's own mirror; the app downloads every file separately and verifies its size and SHA-256.

Attribution

• Original model: alphacep/vosk-model-tg (Vosk Tajik Speech Recognition Model, version 0.61, non-streaming Zipformer2), published by Alpha Cephei under the Apache License 2.0 — a copy is in LICENSE.
• Training data (per the original model card): Peacockery/tajik-asr-corpus-v3, CC BY 4.0 — about 1059 hours of YouTube audio with machine-generated (ElevenLabs) transcripts rather than human ones.
• Changes made (for the Slyshno app): encoder.onnx and joiner.onnx were dynamically quantized to int8 with ONNX Runtime (MatMul nodes, int8 weights) and saved as encoder.int8.onnx and joiner.int8.onnx. decoder.onnx and tokens.txt are the upstream files, unchanged.

Files

file	bytes	sha256
encoder.int8.onnx	70876558	92237a6a51df09a71733e66d25340827ad0a8cba6c3eb6a21ae3b46e77412851
decoder.onnx	2093080	8f0165818d7987052fb1a4396a7523210ac96563fb020e0e6e75aaa2f088475f
joiner.int8.onnx	259417	8a1733b55206775755bc70cee1fb3d9d0a61402042969a0f3d531c9e72b227e8
tokens.txt	5062	1d1c975c225b0f5a994e382d8967e1813d56942651608556283c5aab1f07959f

Usage

sherpa-onnx OfflineRecognizer with a transducer config (model_type = "transducer"), 16 kHz mono input, 80-dim features. The output is lowercase text without punctuation.

Accuracy and speed

FLEURS tg_tj dev, first 12 clips, through sherpa-onnx 1.13.8: int8 WER 16.0 %, CER 5.9 % (fp32: 15.3 % / 5.7 %). The original card reports 13.85 % WER on the FLEURS test split for the fp32 model.

Phones (25 s chunk, 2 threads; real-time factor / peak RSS): Snapdragon 835 0.11 / 367 MB, Kirin 710 0.23 / 367 MB, MT6735m (armeabi-v7a, 2 GB RAM) 0.96 / 282 MB.