TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Cactus Compute announced Whistle, a 16.9 MB speech-recognition model that runs on a CPU and transcribes audio locally in seven languages. The company reports an 11.1 ms time to first token on an Apple M4 Pro for a 10-second clip; benchmark results vary by dataset, and independent verification is not provided in the report.
Cactus Compute announced Whistle on October 2, describing it as a 16.9 MB speech-recognition model that runs on a CPU without dependencies and transcribes speech directly on a device. The company says it supports seven languages and can be loaded in the same C++ engine as its Needle model, a design aimed at mobile devices, wearables, robots, smart-home products, vehicles and microcontrollers.
Whistle handles 16 kHz mono audio clips up to 30 seconds, with English, German, French, Spanish, Italian, Dutch and Polish supported. Cactus says the language is detected automatically unless a user specifies it. Its listed functions include transcription, word-level timestamps with start and end times and probability scores, and speech embeddings that represent the audio without producing a transcript.
The company says audio remains on the device during transcription. Its browser demonstration downloads the model file when a user first presses the microphone button; the report says subsequent processing is local. Whistle also has keyword biasing, intended to raise the likelihood of supplied phrases during decoding. The report describes five-beam search and says clips identified as silence receive an empty transcript without entering that search.
Cactus reports that Whistle takes 11.1 milliseconds to produce its first token for a 10-second audio clip on an Apple M4 Pro CPU. In the same test, it lists 1,319 decoded tokens per second after the first token. The company compares Whistle with Whisper base and Moonshine tiny v2 using each model’s official runtime and default settings. Those results are vendor-reported measurements, not an independent evaluation.
Smaller Models for Local Speech
A model file of 16.9 MB could make speech recognition easier to run in products with limited storage or without a dependable internet connection. If processing stays on-device as Cactus describes, applications may be able to transcribe speech without sending audio to a remote service. That could matter for privacy-sensitive uses and devices that must work offline, though actual privacy protections also depend on how a product stores or handles audio outside the model.
The claimed speed is relevant to interactive applications: a short delay before the first output can make voice interfaces feel more responsive. Cactus’s figures apply to a specified test setup and should not be treated as expected performance on every phone, wearable or embedded processor. The company’s results give an initial basis for comparison, but hardware, runtime settings, audio length and evaluation method all affect performance.
Whistle’s shared engine with Needle is also a product-design choice. Cactus says the models use the same C++ engine, container and quantisation, which may simplify deploying speech recognition alongside Needle. The report says one binary can turn a clip into tool calls, but does not provide a separate evaluation of that broader workflow.
on-device speech recognition software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Whistle Fits Cactus Compute
Cactus presents Whistle as a speech component for its existing Needle CPU engine. The report describes an audio encoder followed by a text decoder: audio is converted into log-mel features, reduced to frames at 80-millisecond intervals, then processed to generate transcript tokens. The model can also return the encoder’s frame-by-frame speech representation without decoding text.
The comparison table gives Whistle a smaller listed file size than the two models tested: Whisper base at 145.3 MB and Moonshine tiny v2 at 41.9 MB. Cactus reports lower word error rates than both competitors on LibriSpeech test-clean and test-other, SPGISpeech, Earnings-22 and the FLEURS average. Its own table also shows Whisper base ahead on TED-LIUM, AMI and the MLS average. Moonshine is listed as English-only, while the report notes that Whisper does not publish results for several datasets in the comparison and that its AMI result uses a different subset.
Those qualifications limit direct comparisons. Word error rate measures transcription errors on a particular dataset; results can vary with the speakers, recording conditions and language represented there. Cactus’s report identifies the test environments and caveats but does not describe an independent replication.
“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 October 2, 2026 announcement
Performance Beyond the Test Setup
The announcement does not provide independent benchmark results or tests across a broad range of phones, microcontrollers and other target devices. Its latency and decoding-speed figures are specific to an Apple M4 Pro CPU, a 10-second clip and the stated runtime setup. Results on lower-powered hardware may differ.
The report does not give enough information to establish how Whistle performs across accents, noisy environments, overlapping speakers or all forms of speech. Its benchmark table covers named datasets, but comparisons are incomplete where competing models lack published results or use different dataset subsets. Cactus also does not identify in the supplied report the licence terms, full model availability details or a release schedule beyond the announcement.
Release Details and Independent Tests
The next useful steps are publication of the model’s access and licensing details, followed by testing outside the company’s reported setup. Developers can compare accuracy, latency, memory use and battery impact on the specific hardware and audio conditions they intend to support. Cactus’s announcement does not state a date for additional benchmarks or a broader release milestone, so those details remain pending.
Key Questions
What is Whistle?
Whistle is a speech-recognition model announced by Cactus Compute. The company says it runs on a CPU and can transcribe audio locally.
Which languages does it support?
Cactus lists English, German, French, Spanish, Italian, Dutch and Polish. The language can be detected automatically or specified by the user, according to the report.
Does audio have to be sent to a server?
Cactus says Whistle processes audio on the device and that audio in its browser demonstration does not leave the device. The report does not establish how every third-party product using the model will handle audio or other data.
How fast is Whistle?
Cactus reports an 11.1 ms time to first token for a 10-second clip on an Apple M4 Pro CPU. That is a company measurement on a specified setup, not a guarantee for other hardware.
How accurate is it compared with other models?
Cactus reports that Whistle leads on several listed datasets, while Whisper base leads on TED-LIUM, AMI and the MLS average. Some comparisons are unavailable or use different dataset subsets, and the announcement does not include independent verification.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
