TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Cactus Compute announced Whistle, a 16.9 MB speech-recognition model designed to run locally on a CPU. The company reports support for seven languages and strong results on several speech benchmarks, but says competing models lead on others; the published speed figures are company tests on an Apple M4 Pro.
Cactus Compute announced Whistle on October 2, describing it as a 16.9 MB speech-recognition model that runs on a CPU without external dependencies. The company says it can transcribe audio locally in English, German, French, Spanish, Italian, Dutch and Polish, a design aimed at devices where network access, latency or sending audio to a cloud service is a concern.
The company says Whistle accepts 16 kHz mono audio clips up to 30 seconds, detects the spoken language unless the user specifies one, and can return word-level timestamps. It also offers speech embeddings: numerical representations of audio produced by the encoder without decoding a transcript. Cactus says audio in its browser demonstration stays on the device after the model is downloaded; that statement describes the demo and should not be read as a guarantee about every application using the model.
Whistle uses the same C++ CPU engine as Needle, another Cactus model, and the company says both can load from the same container and use the same quantisation. Its architecture has an audio encoder and text decoder. Cactus says the encoder processes the clip once, while the decoder uses five-beam search to produce a transcript. Developers can select different decoder depths, but the company says all eight encoder blocks run at each depth.
Cactus reports 11.1 milliseconds to first token for a 10-second clip and 1,319 decoded tokens per second in a test on an Apple M4 Pro CPU. The company compared Whistle with Whisper base and Moonshine tiny v2 using each model’s official runtime and default settings. It says Whistle scored better on LibriSpeech test-clean and test-other, SPGISpeech, Earnings-22 and the FLEURS average. Whisper base scored better on TED-LIUM, AMI and the MLS average. The report notes that several comparisons are unavailable because the other model authors did not publish results for those tests or datasets.
Local Transcription on Small Devices
A 16.9 MB model could make speech recognition more practical on phones, wearables, robots and other hardware with limited storage or connectivity. Running inference locally can also reduce the need to transmit recordings to a remote service. Cactus presents Whistle for these uses, though actual performance and power use will depend on the device, software integration and audio conditions.
The announcement also points to a product-design choice: Whistle shares an engine with Needle, which Cactus says can turn a clip into tool calls in one binary. For developers already using that system, shared runtime components could simplify deployment. That is a claimed integration benefit, not a measured reduction in development time or operating cost.
The reported benchmark results are mixed rather than uniformly better. Whistle’s low file size and speed figures are relevant, but accuracy depends on the dataset and evaluation setup. Readers comparing models should account for differences such as Whisper’s reported AMI-IHM subset versus the AMI dataset used for the other figures.
on-device speech recognition software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Whistle Processes Audio
Cactus describes Whistle as an open speech-recognition model intended for mobile devices, wearables, robots, smart-home products, vehicles and microcontrollers. Its stated input pipeline converts audio into log-mel features, then reduces the representation to 375 frames for a 30-second clip. The company says each remaining frame corresponds to about 80 milliseconds of audio.
The model’s encoder uses eight attention blocks shared with Needle, while its decoder has eight laddered attention blocks and gated cross-attention to the encoded audio. Cactus says the decoder can be run at different depths, with models trained for each supported depth. The release report also describes keyword biasing, word timestamps and a silence check that can return an empty transcript without starting beam search.
In the company’s timing comparison, Whistle’s first-token time rises with clip length: Cactus reports 5.9 milliseconds for five seconds of audio, 11.1 milliseconds for 10 seconds and 36.3 milliseconds for 30 seconds. It says Whisper pads every input to 30 seconds, making that model’s first-token time flat across clip lengths in the test. The report defines decode speed separately from encoder time, so the tokens-per-second figure does not include the audio-processing stage.
“It is one 16.9 MB file, runs on the CPU with no dependencies, and loads into the same C++ engine as Needle.”
— Cactus Compute, in its Whistle announcement
As an affiliate, we earn on qualifying purchases.
Limits of the Published Results
The published figures come from Cactus Compute’s report; the supplied material does not include an independent evaluation or enough detail to reproduce every result. The speed test uses an Apple M4 Pro CPU and each model’s stated official runtime and defaults. Performance on other processors, embedded hardware or real-world recordings may differ.
Accuracy comparisons are also incomplete. Cactus says Moonshine tiny v2 is English-only and that Whisper has no reported figures for several listed benchmarks. Its note that Whisper’s AMI result uses AMI-IHM, a different subset from the AMI results for the other models, further limits direct comparison. The report does not establish how Whistle performs across accents, noisy environments or languages beyond the seven it lists.
The announcement does not specify a detailed licence, hardware requirements beyond CPU operation, power consumption, or a release and maintenance schedule in the supplied material. It is also not clear whether the browser demo’s download size and behavior match all deployment packages. Those details matter for teams evaluating the model for products.
As an affiliate, we earn on qualifying purchases.
Availability and Independent Testing
The next useful milestone is broader testing of the released model on representative devices and recordings. Developers considering Whistle can compare transcription quality, latency, memory use and energy consumption on their own target hardware, while checking how the model handles their required languages and acoustic conditions.
Cactus has announced the model and provided a browser demonstration, but the supplied report does not give a date for further releases, independent benchmark results or additional language support. Until those details are available, the clearest confirmed picture is the company’s stated model size and feature set, alongside benchmark and speed claims tied to its described test setup.
speech recognition for small devices
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Whistle?
Whistle is a speech-recognition model announced by Cactus Compute. The company says it runs on a CPU and transcribes audio locally.
Which languages does Whistle support?
Cactus lists English, German, French, Spanish, Italian, Dutch and Polish. It says the model detects the language unless the user specifies one.
Does Whistle send recordings to the cloud?
Cactus says audio in its browser demonstration stays on the device after the 16.9 MB model download. The announcement does not establish the data handling of every third-party application built with Whistle.
How does Whistle compare with Whisper?
Cactus reports that Whistle is smaller and faster in its Apple M4 Pro test, and leads on some listed accuracy benchmarks. It says Whisper base performs better on TED-LIUM, AMI and the MLS average, so results depend on the test.
What remains unknown about Whistle?
The supplied report does not provide independent verification, detailed licensing and hardware requirements, power-use measurements, or results across all accents and noisy conditions. Performance on devices other than the test system also remains to be established.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
