v0.3.2

Contributors:
@ZachNagengast @keith4ever @dylanangus @chen-argmax @flashno @bpkeene

What's Changed

From v0.3.0:

• Performance improvements
• Imports the standard HuggingFace tokenizer via C bindings from https://github.com/FL33TW00D/tokenizers-sys
• Updated example app
  • Automatic model downloads from HuggingFace
  • Transcribe from file or recording from the microphone
  • Selectable compute units between NPU, GPU, and CPU
  • Thank you to @Acs176 for kickstarting this initiative
• Publishable Maven repository for Kotlin library
• Code style rules with auto-formatting via make format

with v0.3.2:

• Add multilingual model support with PerLayerKVDecoder, handling diverse encoder output names, int64 token bindings, and timestamp postprocessing; lifts KV-cache logic into decoder for improved modularity.

• Various SDK and build improvements: conditional Bazel download, updated API with segments, improved model download/handling, version bumps (0.3.2), and README cleanup.

• Example app published for open testing: https://play.google.com/store/apps/details?id=com.argmaxinc.whisperax

  • Demo:
    

v0.1.0

Beta release for WhisperKitAndroid!

What's Changed

• Added Linux support (CPU only) for model pipeline execution
• Testing coverage across earnings-22, librispeech datasets
• Initial C-API and internal runtime
• whisperkit-cli command line utility, with subset of features as from WhisperKit
• Reduced exported symbols, improved build system, and other improvements