Whistle publishes a 16.9 MB seven-language model and CPU runtime for transcription, timestamps and speech embeddings.
Whistle runs multilingual speech recognition in a 16.9 MB edge model
2026-10-01Verified announcementTRIAL
Cactus Compute released Whistle, a 16.9 MB Apache-2.0 speech model for seven languages. The same CPU runtime can return transcripts, word timestamps and frame-level speech embeddings without a GPU, with Python, C, platform binaries and WebAssembly deployment paths. Benchmarks are author-reported.
What this means for you
For teams that need private or offline speech features on constrained devices, Whistle combines transcription and embeddings in one small artifact with an inspectable runtime. Trial it on real accents, noise, silence, keyword bias and target-device thermals before replacing established ASR.
Evidence & related material
Technical details & scores
| Size | 16.9 MB model artifact |
|---|---|
| Languages | English · German · French · Spanish · Italian · Dutch · Polish |
| Outputs | transcription, word timestamps with probabilities, and 80 ms frame speech embeddings |
| Runtime | CPU-only C++ runtime with Python, C, platform binaries and WebAssembly paths |
| Reported Evaluation | WER over 86,174 utterances; comparisons use author-published Whisper and Moonshine figures |
| Community Signal | 2,594 downloads and 161 likes at review time |
Technical impact 3
Engineering value 4
Adoption signal 3
Maturity 3
Verification cost 2
Risk 2