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 device’s CPU and supports seven languages. The company reports low latency and competitive results on several speech benchmarks, but Whistle does not lead every comparison and the figures come from tests run with different models’ official runtimes.
Cactus Compute announced Whistle, a speech-recognition model packaged in a 16.9 MB file that runs on a device’s CPU without external dependencies. The company says it can transcribe audio in seven languages, return word-level timestamps and produce speech embeddings, targeting uses such as mobile devices, wearables, robots and cars.
Whistle handles 16 kHz mono audio clips of up to 30 seconds in English, German, French, Spanish, Italian, Dutch and Polish. Cactus Compute says the language is detected automatically unless the user specifies it. The company’s in-browser demonstration downloads the model on first use and says the audio remains on the device; the report describes local processing as a core feature.
Alongside transcripts, Whistle can provide start and end times for each word, with a probability value, or return speech embeddings without decoding a transcript. The model includes keyword biasing, intended to raise the likelihood of supplied phrases in the output. A silence check can return an empty transcript without running beam search, according to the technical description.
Cactus Compute reports that Whistle reaches its first token in 11.1 milliseconds and decodes at 1,319 tokens per second on an Apple M4 Pro CPU using a 10-second clip. These are company-reported measurements: its comparison used each model’s official runtime and default settings, including five beams for Whistle. The report says Whistle’s time to first token varies with clip length, measuring 5.9 ms at five seconds and 36.3 ms at 30 seconds.
Local Speech Processing on Small Devices
A 16.9 MB model could make speech recognition easier to run on devices with limited storage or without a reliable internet connection. Processing speech locally may also reduce the need to send recordings to a remote service, though privacy depends on how a particular product is built and configured. Cactus Compute’s browser demo is described as keeping audio on the device.
The potential audience ranges from wearables and smart-home equipment to robots and automotive systems, where response time, connectivity and computing limits can shape which speech tools are practical. A small download and reported fast response may help developers test local transcription in those settings. The announcement does not establish how Whistle performs across all devices, under sustained workloads or in real-world noisy environments.
The benchmark results also point to a trade-off rather than a universal win. Cactus Compute says Whistle scores ahead of Whisper base on several listed datasets, while Whisper base leads on TED-LIUM, AMI and the MLS average. Readers comparing models should account for dataset coverage, model size and test setup, rather than treating a single speed or accuracy figure as decisive.
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’s Models
Cactus Compute describes Whistle as sharing the same C++ CPU engine and model components as its Needle model. The company says both load from the same container and use the same quantisation; blocks identified as shared use Needle’s code. Whistle adds speech-specific components, including a decoder that attends to the audio representation while producing text.
The speech pipeline converts audio into log-mel features, then reduces the representation to 375 frames for a 30-second clip. The encoder processes those frames, while the decoder generates text using beam search. The published report says users can select decoder depth when loading the model, but the encoder still runs all eight of its blocks.
In its size comparison, Cactus Compute lists Whisper base at 145.3 MB and Moonshine tiny v2 at 41.9 MB, alongside Whistle at 16.9 MB. The company’s benchmark notes caution that coverage differs: Moonshine is English-only, and some systems lack published results for particular datasets. It also says Whisper’s AMI score uses a different subset from the AMI results reported for the other two models.
““Speech recognition” on the device, in “one 16.9 MB file.””
— Cactus Compute
As an affiliate, we earn on qualifying purchases.
Benchmark Coverage and Device Performance
The announcement does not provide independent testing or enough detail to confirm how the reported speed and accuracy figures generalize beyond the stated setup. Its latency and decoding measurements use an Apple M4 Pro CPU, a 10-second clip and different official runtimes; performance on phones, microcontrollers and other supported target devices is not given in the source material.
Accuracy results also have limits. Cactus Compute reports that Whistle leads on LibriSpeech test-clean and test-other, SPGISpeech, Earnings-22 and the FLEURS average, while Whisper base leads on TED-LIUM, AMI and the MLS average. The report notes missing published results and differences in the AMI subset. It does not provide enough information here to assess all test conditions, language-by-language performance, or how the model handles accents, background noise and overlapping speech.
The announcement identifies intended uses across several device categories, but does not specify licensing terms, minimum memory requirements, power consumption or a full compatibility list. Those details affect whether developers can deploy the model in a particular product.
As an affiliate, we earn on qualifying purchases.
Deployment Details to Watch
The next practical test is whether Whistle’s compact size and reported speed hold up on the devices Cactus Compute names, including phones, wearables and embedded systems. Developers will also need technical details such as supported platforms, memory use, power draw and licensing before judging whether it fits a production application.
Further independent evaluations could clarify accuracy across all seven languages and compare results under consistent hardware, audio and runtime conditions. Until those details are available, the launch establishes the model’s announced capabilities and company-reported measurements, not a settled ranking against other speech-recognition systems.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Whistle?
Whistle is a speech-recognition model released by Cactus Compute. It is designed to run locally on a device’s CPU and is distributed as a 16.9 MB file, according to the company.
Which languages does Whistle support?
The announcement lists English, German, French, Spanish, Italian, Dutch and Polish. Cactus Compute says Whistle detects the language unless the user names it.
Does Whistle send audio to a server?
Cactus Compute says Whistle runs on-device and its browser demo keeps audio on the device. The announcement does not document how every possible third-party product using the model will handle recordings.
Is Whistle more accurate than Whisper?
Not across every benchmark listed. Cactus Compute reports better Whistle results on several datasets, while Whisper base leads on TED-LIUM, AMI and the MLS average. Dataset coverage and AMI subsets differ, so the comparison is not a single overall ranking.
How fast is Whistle?
Cactus Compute reports 11.1 ms to first token and 1,319 tokens per second for a 10-second clip on an Apple M4 Pro CPU. These are company-reported results from the stated test setup, not a guarantee of performance on other devices.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
