unstale

npm / sherpa-onnx-node

sherpa-onnx-node

What changed in sherpa-onnx-node 2, from its official release notes, upgrade guides and advisories. Sources checked 2026-11-04.

1 breaking1 security8 new APIs and options1 behavior and defaults15 fixes and other

See which of these apply to your code, ranked, in your agent:

npx @unstale/cli init

@unstale/cli on npm

sherpa-onnx-node 226 changes

Breaking changes 1

  • Raises ValueError when it cannot encode keywords.

    2.0.0release notes

Security 1

  • Fixed a heap overflow for short text when speed is greater than 1.

    2.0.0release notes

New APIs and options 8

  • FunASR-Nano prompt options can now be set per stream.

    2.0.0release notes

  • Added GetEmbedding to SpeakerEmbeddingManager.

    2.0.0release notes

  • Offline CTC models now support HLG decoding through the C, CXX, JNI/Kotlin/Java, and WASM APIs.

    2.0.0release notes

  • Added this option to online-websocket-server.

    2.0.0release notes

  • Streaming NeMo transducer models now support hotwords with modified_beam_search.

    2.0.0release notes

  • Added OfflineDiacritization to the Kotlin API.

    2.0.0release notes

  • A Node.js addon is now available for Windows arm64.

    2.0.0release notes

  • The Flutter package now ships arm64 libraries.

    2.0.0release notes

Behavior and default changes 1

  • Provider configuration keys are now forwarded to the CUDA execution provider.

    2.0.0release notes

Fixes and other changes 15

  • SherpaOnnxCreate* functions catch C++ exceptions and return NULL.

    2.0.0release notes

  • lang2id is derived from the model vocabulary, and the system warns instead of silently falling back to English.

    2.0.0release notes

  • Fixed strings returned to Java for characters outside the BMP.

    2.0.0release notes

  • C++ exceptions are converted to Java exceptions in these constructors.

    2.0.0release notes

  • Fixed word boundaries, out-of-vocabulary words, and Unicode handling.

    2.0.0release notes

  • Fixed decoding by feeding encoder_attention_mask at raw audio length.

    2.0.0release notes

  • NeMo mel features now use NeMo's additive log-zero guard.

    2.0.0release notes

  • Fixed timestamps after filtering special tokens.

    2.0.0release notes

  • Fixed an abort when input contains fewer than 400 samples.

    2.0.0release notes

  • Fixed JSON escaping in online and offline recognition results.

    2.0.0release notes

  • Python methods now check embedding dimensions before releasing the GIL.

    2.0.0release notes

  • Python methods check the embedding size.

    2.0.0release notes

  • The language-agnostic na tag is now accepted.

    2.0.0release notes

  • VAD sample positions were widened to int64_t to fix SIGSEGV on long streams.

    2.0.0release notes

  • Fixed character loss for byte-level BPE tokens.

    2.0.0release notes