Whistle: a 16.9MB speech-to-text model that beats Whisper base at 9x smaller size and 6x speed
Henrie_the_dreamer · reddit · 2026-10-06
Cactus Compute released Cactus Whistle, an ASR model for ultra-small devices: 55M params (36M active), quantized to a 16.9MB file — 9x smaller than Whisper base (145.3MB) and 6x faster, while mostly beating it: 4.31 WER on LibriSpeech test-clean (vs 4.9), 10.49 on test-other, 21.4 FLEURS average (vs 24.5). It supports English, German, French, Spanish, Italian, Dutch, and Polish.
Architecture: log-mel front end + conv stem into an audio encoder; a Simple Attention + Hadamard MLP decoder with gated cross-attention, ladder-structured so any depth from 2 layers up is deployable. Keyword biasing favors user-specific names during beam search, and word timestamps come from decoder attention. It ships on 17 platforms including RISC-V, WebAssembly, and watchOS — the goal is compressing intelligence onto budget phones, wearables, and microcontrollers.
Related event: Tiny ASR Model Whistle Runs in Browsers at 16.9MB(2 posts)→
More from Infra
- Auction-based LLM routing paper finds fixed token pricing wastes 10%-71% in costs — _onionesque · 2026-10-18
- Running Ling 3.1 Flash (560B MoE, 25B Active) Across 4 DGX Sparks via llama.cpp RPC — BerryBlush_Rogers691 · 2026-10-18
- SGLang Summit sponsors jump from 11 to 36 in one week — ying11231 · 2026-10-18
- Oracle CDS hits record 300bps, called first AI trade domino after Barclays sell downgrade — arthurcolle · 2026-10-18
- Agent Coding Workflow Would Cost ~$40k/Month on API Pricing, Dev Reveals — ChrisRackauckas · 2026-10-18
- PS5LM: llama.cpp port turns a jailbroken PS5 into a home LLM server (500+ GitHub stars) — wateriscoding · 2026-10-18