Whistle: Speech to Text in 16.9 MB
- ID
- 33447
- Status
- summarized
- Published
- 09 Oct 2026, 12:59 AM
- Fetched
- 09 Oct 2026, 6:35 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://cactuscompute.com/blog/whistle
- Source URL
- https://hnrss.org/best
Summary
- Score
- 8.0
- Created
- 09 Oct 2026, 6:36 AM
- Tags
- Audience
- developersvibe_codersai_ml_learnersai_agent_userssaas_founders
What happened
Whistle is a 16.9 MB open speech-recognition model from CactusCompute that runs on CPU with no dependencies, shares Needle's C++ engine and quantisation, and transcribes up to 30 seconds of 16 kHz mono audio in English, German, French, Spanish, Italian, Dutch or Polish. It returns word timestamps and speech embeddings (one row per 80 ms frame), claims 11 ms to first token, and its sandbox runs in-browser with audio not leaving the device. The HN discussion has 408 points and 99 comments; the post gives architecture details but not license, accuracy benchmarks, or production support.
Why it matters
For builders adding voice, this is a concrete alternative to cloud STT: a 16.9 MB CPU model could ship inside mobile, wearable, or MCU apps, with timestamps and embeddings for agent tool calls. For Malaysia/Southeast Asia-facing products, the lack of Malay, Mandarin, Tamil, or Manglish support means you would likely need fine-tuning or a hybrid approach before relying on it; verify license and real accuracy before replacing paid APIs.
Discussion angle
Would you use Whistle for a local voice agent despite no Malay/Manglish support, and what would it take to fine-tune or add those languages on top of the shared Needle engine?