Whistle Explained: Cactus's 16.9MB Speech-to-Text Model

Whistle is an open speech-to-text (ASR) model released by Cactus Compute on October 2, 2026. Its defining trait is that the whole model is a single 16.9MB file that runs on a CPU with no dependencies. It is licensed under Apache 2.0, and the parameter count has not been disclosed. It runs on the same C++ engine, the same .cact container format, and the same quantization as the company's tiny agent LLM (see our Needle 2 explainer), so you can chain speech recognition to tool calling in a very small footprint. It hit #1 on the Hacker News front page on October 8, 2026, with about 768 points.

Important note: Whistle does not support Japanese or Chinese. It covers seven languages: English, German, French, Spanish, Italian, Dutch, and Polish. If you need Japanese transcription, choose a Whisper-family or other Japanese-capable model (see the comparison table below).

- Transcription: processes up to 30 seconds of 16 kHz mono audio per pass; language is auto-detected or set explicitly, e.g. language="de"
- Word timestamps: start, end, and probability for each word
- Speech embeddings: one encoder row per 80 ms frame, available without decoding
- Keyword biasing: nudge the model toward names or command words you specify
- Silence detection: returns an empty transcript before decoding when it detects silence

Silence detection helps with the classic problem of phantom text appearing over quiet audio. As covered below, though, community members still reported false output on TV audio.

Input audio becomes 80 log-mel bins (25 ms window, 10 ms hop, 250–3500 Hz), and a convolutional stem compresses 3000 frames down to 375. The encoder is 8 non-causal Simple Attention blocks. The decoder is 8 Laddered Simple Attention blocks of width 512, with GQA (8 query heads, 2 KV heads), engram lookups at layers 3 and 7, and gated cross-attention in every decoder layer. Decoding uses beam search with 5 beams and up to 320 tokens; the vocabulary is 8,192 plus 7 language tokens. The --audio-depth option selects how many decoder layers to use at load time (minimum 2).

All figures are vendor-reported. Whistle's TTFT is reported as 5.9 ms on a 5 s clip and 36.3 ms on a 30 s clip. For accuracy, the vendor compared WER across 86,174 utterances: Whistle comes out ahead of Whisper base and Moonshine tiny v2 on LibriSpeech clean/other, SPGISpeech, Earnings-22, and the FLEURS average, while Whisper base is ahead on TED-LIUM, AMI, and the MLS average. Exact numbers appear only in the vendor's chart, so we describe the trend only.

Prebuilt engines are provided for 17 targets, including macOS, Linux, Android, iOS, watchOS, Windows on ARM, RISC-V, MIPS, the browser, and WASI; each ships a needle binary, libneedle.a, and needle.h. Weights are on Hugging Face, the engine is at the needle3 repo, and the source is on GitHub.

From Python, pip is the fastest route.

For microphone capture or audio that is not 16 kHz, install the mic extra (soxr, sounddevice).

In Python, transcription is one line. Options cover word timestamps, keywords, and language.

The CLI covers downloading weights, transcribing files, trying the mic, and comparing against Whisper and Moonshine.

To go straight from voice to tool calls, pass both Needle and Whistle models together.

For native deployment, fetch the platform engine first, then Whistle.

Whistle's strengths come down to size and first-response speed. For general-purpose transcription accuracy, larger models still have the edge. If you want local Japanese transcription, a Whisper-based app such as the one in our TypeWhisper explainer is the realistic choice. And if you also want Japanese speech output after recognition, see the Irodori-TTS v4 explainer.

The Hacker News thread mixed surprise at the size and speed with sober criticism. Many commenters felt raw accuracy falls short of Whisper Large, Whisper large-v3-turbo, and Parakeet. Hallucinations on silence and TV audio were reported, and the public demo has no streaming output. Some also said the name collides with an existing Android app called Whistle, which caused confusion.

Narrower use cases got a different verdict. One user tested 170 command phrases: raw Whistle got 70 right, versus 168 for Qwen ASR 1.7B. After training a small template-restricting network on 10,000 generated utterances, Whistle reached 164, and on another phrase set it scored 206 of 208 against 184 for Vosk small. The takeaway is that the strongest use case is fixed-command voice control on tiny devices (for example, Home Assistant on an Echo Show), not general dictation.

- Fits: fixed-command recognition in English and European languages; voice control on wearables, robots, automotive, and microcontrollers; setups that go directly from voice to tool calls
- Doesn't fit: Japanese or Chinese transcription; long, high-accuracy jobs such as meeting minutes; anything that requires streaming output

If Japanese voice input is your goal, pick a Whisper-family model or one that explicitly lists Japanese support. Consider Whistle only for on-device command recognition in English and European languages.