Starstruk Logo MarkStarstruk

v1.12.18

k2-fsa/sherpa-onnxv1.12.18Nov 27, 2025by csukuangfj

AI Summary

Version 1.12.18 expands NPU support significantly, adding Qualcomm QNN and Huawei Ascend NPU support for various models like SenseVoice, Zipformer CTC, and Paraformer. It also exports the OmniLingual ASR CTC 1B model and fixes a crash when reading non-wav files.

Key Highlights

  • Added C++ QNN support for SenseVoice and Zipformer CTC models.
  • Added Android demo with QNN for SenseVoice ASR.
  • Added C++ support for Zipformer CTC and Paraformer on Ascend NPU.
  • Exported OmniLingual ASR CTC 1B model.
  • Fixed segfault when non-wav file is passed to ReadWave.

New Features

  • QNN support (SenseVoice/Zipformer)
  • Ascend NPU support (Zipformer CTC/Paraformer)
  • Omnilingual ASR CTC 1B export
  • Hotwords (byte level BPE)
  • Symbol visibility limiting
  • Whisper Large V3 external data support

Full Release Notes

## What's Changed
* Fix CI by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2786
* export omniASR_CTC_1B from https://github.com/facebookresearch/omnilingual-asr by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2788
* Add C++ QNN support for SenseVoice by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2793
* Export models for CANN toolkit 7.0 by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2795
* Support hotwords with byte level bpe by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2802
* Add Android demo with QNN (Qualcomm NPU) for SenseVoice ASR by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2803
* Export zipformer ctc models to QNN by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2815
* Add spaces between English words for Homophone replacer. by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2817
* Add C++ QNN support for Zipformer CTC models. by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2809
* Limit symbol visibility in the shared libraries by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2822
* Fix warnings for initializing tts lexicon. by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2823
* Export zipformer ctc models to Ascend NPU by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2824
* Refactor scripts for exporting models to Ascend NPU. by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2825
* Add C++ support for Zipformer CTC on Ascend NPU by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2826
* Fix segfault when non-wav file is passed to ReadWave by @jcheng5 in https://github.com/k2-fsa/sherpa-onnx/pull/2821
* Avoid calling rknn_dup_context(). by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2828
* Add C++ support for Paraformer with RK NPU by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2829
* Update README to include NPU support by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2830
* Support running whisper large v3 with external data weight by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2807
* Release v1.12.18 by @csukuangfj in https://github.com/k2-fsa/sherpa-onnx/pull/2831

## New Contributors
* @jcheng5 made their first contribution in https://github.com/k2-fsa/sherpa-onnx/pull/2821

**Full Changelog**: https://github.com/k2-fsa/sherpa-onnx/compare/v1.12.17...v1.12.18