What is Cactus Whistle?

Whistle is an open speech-recognition model released by Cactus Compute on October 2, 2026. It is a single 16.9 MB file that runs on CPU with no dependencies, transcribes up to 30 seconds of 16 kHz mono audio per pass in English, German, French, Spanish, Italian, Dutch and Polish, returns word timestamps with probabilities, and loads into the same C++ engine as Cactus's Needle tool-calling model so one binary can turn a voice clip straight into function calls.

Is Whistle more accurate than Whisper Base?

On some benchmarks. Cactus reports Whistle ahead of Whisper Base on LibriSpeech test-clean and test-other, SPGISpeech, Earnings-22 and the FLEURS multilingual average, while Whisper Base stays ahead on TED-LIUM, AMI and the MLS average. Whistle's numbers come from Cactus's own runs over 86,174 utterances; Whisper's are the figures OpenAI published for the multilingual checkpoint, so the comparison is vendor-run but uses standard Whisper normalizers.

How fast is Whistle compared to Whisper Base and Moonshine?

On an Apple M4 Pro CPU with ten seconds of audio, Cactus measured Whistle at 11.1 ms time to first token and 1,319 tokens per second decode, versus 73.2 ms and 266 tokens/s for Whisper Base and 22.8 ms and 262 tokens/s for Moonshine Tiny v2. Whisper pads every input to 30 seconds so its latency is flat; Whistle's scales with clip length, from 5.9 ms at 5 seconds to 36.3 ms at 30 seconds.

Which tiny speech-to-text model should I use on a microcontroller or wearable?

Whistle if you need any of its seven languages, word timestamps, keyword biasing or a combined speech-plus-tool-calling pipeline, and your clips are 30 seconds or shorter. Moonshine Tiny v2 (41.9 MB) if you need English-only true streaming with sub-200 ms latency on devices like a Raspberry Pi. Whisper Base (145.3 MB) if you need any of Whisper's 99 languages or the mature whisper.cpp ecosystem and can afford roughly nine times the storage.

Quick Answer

Whistle vs Whisper Base vs Moonshine Tiny: Tiny STT (2026)

Published: October 4, 2026

The short answer

Cactus Whistle (October 2, 2026) is the smallest credible on-device speech-to-text model available: 16.9 MB, 11.1 ms to first token, and ahead of the 145.3 MB Whisper Base on LibriSpeech, SPGISpeech, Earnings-22 and FLEURS. Whisper Base still wins on TED-LIUM, AMI and MLS and covers 99 languages to Whistle’s seven. Moonshine Tiny v2 sits between them at 41.9 MB and is the only one of the three built for genuine streaming. Facts verified October 4, 2026 from the Cactus release post and the models’ published cards.

Side by side

	Whistle	Whisper Base	Moonshine Tiny v2
Maker	Cactus Compute	OpenAI	Useful Sensors
Released	Oct 2, 2026	Sep 2022 (base)	2025 (v2)
Size on disk	16.9 MB	145.3 MB	41.9 MB
Languages	7 (en, de, fr, es, it, nl, pl)	99	English only
Max clip per pass	30 s	30 s (padded)	Streaming
Time to first token (10 s clip, M4 Pro CPU)	11.1 ms	73.2 ms	22.8 ms
Decode speed	1,319 tok/s	266 tok/s	262 tok/s
Word timestamps	✅ with probability	Via tooling	✅
Keyword biasing	✅ (Aho-Corasick)	❌	❌
Dependencies	None (single C++ engine)	PyTorch / whisper.cpp	ONNX / own runtime
Licence	Open (see HF card)	MIT	MIT

Speed figures are Cactus’s, measured on an Apple M4 Pro CPU with each model on its official runtime at defaults: Whistle’s engine at 5 beams, openai-whisper, and moonshine-voice non-streaming over whole audio.

What makes Whistle 16.9 MB

Whistle is not a distilled Whisper. It reuses the architecture of Cactus’s Needle tool-calling model: eight Simple Attention encoder blocks (non-causal, so a frame at 3 s can attend to a frame at 12 s) and eight laddered decoder blocks at width 512 with 8 query heads to 2 KV heads. The only speech-specific addition is a gated cross-attention per decoder layer that reads the encoder output. Those cross-attention projections run once when the clip arrives and are then held for the whole decode, so five-beam search costs five short transcript caches rather than five passes over the audio.

Two engineering choices matter for embedded use:

The decoder is a ladder. Every depth from 2 layers up was trained as its own model, and --audio-depth picks one at load time. You trade accuracy for latency without downloading a different file. The encoder always runs all eight blocks.
Silence short-circuits. The engine measures loudness range before decoding and returns an empty transcript below threshold, never entering beam search. On a wake-word device that is most of the calls.

The vocabulary is 8,192 text pieces plus seven language tokens, so the detected language comes back as a token rather than a side channel. Transcripts are capped at 320 tokens.

Accuracy, honestly

Whistle’s WER is measured by Cactus over 86,174 utterances with Whisper normalizers; Whisper’s and Moonshine’s are the figures their authors published for multilingual checkpoints. Cactus verified by audio checksum and speaker ID that no test audio leaked into training. With that caveat:

Whistle ahead: LibriSpeech test-clean and test-other, SPGISpeech, Earnings-22, FLEURS average.
Whisper Base ahead: TED-LIUM, AMI, MLS average. (Whisper’s AMI number is AMI-IHM, a different subset.)
Moonshine: English only, so it has no FLEURS or MLS numbers to compare.

The pattern is consistent with the design: Whistle is strong on read and business speech and on the seven languages it targets, weaker on long-form lecture audio where Whisper’s bigger decoder helps. If your corpus looks like TED talks, Whisper Base’s extra 128 MB buys real accuracy. If it looks like voice commands and earnings calls, it does not.

Speech straight into tool calls

The feature that separates Whistle from every other tiny ASR model is the engine, not the weights. needle_load reads whichever model a .cact file holds, so:

needle --model needle3.cact --model whistle.cact --tools tools.json --audio clip.wav

returns one JSON object with both the function calls and the speech fields (audio_text, audio_language) — the caller never touches a transcript. In Python it is pip install cactus-needle and needle.transcribe("clip.wav")["text"]; keywords=["Siobhan", "Krzysztof"] biases the beam search toward names your users actually say. For a smart-home or automotive voice stack that is the difference between two models with glue code and one binary.

Which to pick

Microcontroller, wearable or any device where 16.9 MB vs 145 MB is the decision: Whistle.
Voice commands feeding a local tool-calling agent: Whistle + Needle, one runtime.
English-only live captions with true streaming: Moonshine Tiny v2. Whistle and Whisper both work in 30-second passes.
Any language outside Whistle’s seven, or lecture/long-form audio: Whisper Base via whisper.cpp — or step up to a larger local model if the device allows.

Last verified: October 4, 2026. Benchmark and latency figures are Cactus Compute’s published results; Whisper and Moonshine WER figures are from their authors.