Whistle: Speech to Text in 16.9 MB
A new 16.9 MB speech‑to‑text model, Whistle, is drawing attention for delivering surprisingly strong, fully on‑device transcription—especially for English—within a footprint small enough to embed in keyboards, web apps, and low‑power hardware. Commenters compare it with larger systems like Whisper and Parakeet, praising its speed, privacy and deployability while noting weaker performance for some accents and non‑English languages, limited streaming UX, and higher error rates in more demanding scenarios. The thread highlights a broader shift toward small, CPU‑friendly ASR models tailored for dictation, offline assistants, and constrained devices, even as accuracy trade‑offs remain a key concern.
Overall reception
- Many are impressed that a 16.9 MB, CPU-only model achieves strong English transcription, sometimes “insanely” good for normal speech and tech terms.
- Others report noticeably high error rates, comparing quality to older dictation systems and saying size is irrelevant if accuracy is poor.
- Perceived quality varies heavily by accent, recording conditions, and language.
Language & accent performance
- English generally fares best, including for some non-native or heavy accents, though some users still see dropped words or mis-hearings.
- Spanish feedback is mixed: Castilian can work if spoken slowly and clearly; Mexican/Venezuelan users see bad errors and invented words.
- Reports of poor or “meh” results for Polish, Spanish, and French unless speech is very clear.
- Users note lack of support for languages like Japanese, Norwegian, and Chinese; attempts fall back to wrong languages or nonsense.
Model size and on-device benefits
- Small size is praised for enabling:
- On-device, private STT on phones, keyboards, embedded hardware, and webapps.
- Potential to fit in CPU cache, reducing memory latency and enabling “always on” use.
- Some question focusing on size when larger models (Parakeet, Whisper, Qwen ASR) are more accurate.
Comparisons to other STT systems
- Parakeet is frequently cited as a strong, fast local baseline, with several saying it outperforms Whistle and some saying it’s “gold standard.”
- Others note Whisper Large(-v3/-turbo), Qwen3-ASR, Canary, Deepgram, Nemotron, and various apps (Handy, FUTO keyboard, ZWhispr, etc.) as more accurate but heavier or cloud-based.
- Some argue that beyond a certain point, residual STT errors can be repaired by a downstream LLM, especially for notes and dictation.
UX, streaming, and grammar-based setups
- Multiple commenters insist real-time streaming output is essential; the demo’s record-then-transcribe flow is seen as a major gap.
- Users discuss custom dictation workflows (push-to-talk, clipboard paste, delayed “flush”), and ask how streaming handles later corrections.
- One detailed example uses Whistle only as a front-end to a small grammar/template classifier on a rooted smart display, yielding near–SOTA accuracy for a limited command set.
Ethics and AI-assisted cleanup
- Several users describe a common workflow: record audio → STT → LLM cleanup/structuring.
- One commenter worries such AI-polished writing will be dismissed as “AI slop”; others respond that light cleanup on human-generated text is ethically fine and likely overthought.