Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Whistle packs speech recognition into a 16.9 MB file for phones and robots

Cactus Compute on 3 October released Whistle, a speech recognition model that ships as a single 16.9 MB file and runs on a CPU without dependencies, transcribing seven languages. It reaches the first token in 11 ms, the company says.

AI & modelsExplainerGrace OkonkwoPublished: 3 October 20263 min readSources 8
Whistle packs speech recognition into a 16.9 MB file for phones and robots

Cactus Compute published Whistle on 3 October, describing it as a model aimed at mobiles, wearables, robots, smart home devices, automotive systems and microcontrollers. According to the company's blog post, the whole model is one 16.9 MB file that loads into the same C++ engine as its earlier Needle model, from the same container and the same quantisation.

Why the size matters: Whisper base, the comparison Cactus uses, is 145.3 MB. That is roughly 8.6 times the footprint. For anyone building a voice interface on a device with little memory or no reliable connection, the difference decides whether the model fits at all. What Whistle does not settle is accuracy on messy, real-world audio: Cactus reports wins on some benchmarks and losses on others, and no independent replication is public yet.

On the numbers Cactus published, Whistle leads on LibriSpeech test-clean and test-other, on SPGISpeech, on Earnings-22 and on the FLEURS average. Whisper base leads on TED-LIUM, on AMI and on the MLS average. Each model ran on its official runtime at its defaults, with Whistle's engine at five beams, according to the post.

The language list is English, German, French, Spanish, Italian, Dutch and Polish. The model detects the language unless you name it, and emits the detected language as a token in the vocabulary of 8,192 text pieces plus seven language tokens. It handles 16 kHz mono audio of any length, one pass up to 30 seconds and windowed beyond it.

What else landed in on-device speech

Whistle is not the only small model to arrive this week. Shield, from ZeroLeaks, is a 118M-parameter classifier for prompt injection and jailbreak detection that runs locally on a CPU with 118 MB of int8 ONNX weights. Its release S15e is dated 2 October. It does not transcribe speech, but it occupies the same slot: a small model that checks text before an agent acts on it, offline. The model card reports 84.3% mean category-balanced accuracy and 71.8% attack recall with Shield rules v2, measured on 35 public datasets that informed development rather than an unseen test.

Aleph Alpha also published weights this week. Kolibri, released on 3 October, is an English-German mixture-of-experts transformer with 78B total parameters and 3B active, a context window of up to 1M tokens and an Apache 2.0 licence. The company says it matches models with up to four times its active parameter count, naming Nemotron 3 Super. It is a text model, not a speech one, and Aleph Alpha positions it for public administration, industrials and aerospace.

The clearest sign of where speech tooling is heading comes from a much smaller project. On 3 October, Kitsuno published an engineering write-up on training an open decision model, Laya, to stand in for a closed one, Jev, in production. Laya reads text and returns typed decisions with probabilities. Kitsuno reports that on a staging test it found the right country for 26% of job ads and the right work mode for 50%, against 89% for Jev.

The company says it trained on none of Jev's answers, citing a customer agreement clause that bars using output to train a competing product.

That detail tracks a wider shift. The models getting released now are narrow, small and meant to run on hardware you own, whether that is a phone, a robot or a server in Frankfurt. Whistle's 11 ms first-token figure and its 16.9 MB file are the pitch. Whether the accuracy holds outside a benchmark harness is the open question.

Cactus has not said when Whistle will appear in a shipping product, and neither has any device maker. The weights and the engine are the release, for now.

Comments 0

Sources

8
  1. 01Whistle: Speech to Text in 16.9 MBEN
  2. 02Shield: A 118M model for detecting prompt injections and jailbreaksEN
  3. 03Kolibri: A Sovereign Open-Weight ModelEN
  4. 04Training an open decision model to replace a closed one, in shadow on prodEN
  5. 05New in Llama.cpp: Decision ModelsEN
  6. 06Anthropic's super bug-hunting model Mythos is hardcore good at math, as latest vuln under attack showsEN
  7. 07Anthropic claims popular Chinese AI model has Mythos-class hacking abilitiesEN
  8. 08A Beginner's Guide to Running AI Models LocallyEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Grace Okonkwo

Grace Okonkwo

AI, models and technology

Grace Okonkwo covers AI, models and technology for FLASH24, working from primary sources such as model cards, API documentation and benchmark papers rather than vendor summaries. She checks training data provenance, evaluation conditions and reported scores against the underlying datasets before any figure reaches print. She interviews researchers and engineers directly, tracks release calendars from major labs, and compares successive model versions on the same tests. Her own self-hosting, home-network and documentation-reading habits feed straight into that desk, since she tests tools on her own hardware first. She does not publish benchmark claims without a reproducible method.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.