TL;DR
Get business pricing on networking and server gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Cactus Compute announced Whistle, a 16.9 MB open speech-recognition model designed to run locally on a CPU. The company says it supports seven languages, word-level timestamps and speech embeddings; its published benchmarks show mixed results against larger models, with rankings varying by dataset.
Cactus Compute announced Whistle on October 2, describing it as an open speech-recognition model packaged in a 16.9 MB file and designed to run locally on a CPU. The company says it can transcribe speech in seven languages, return word timestamps and produce speech embeddings, a combination aimed at devices where connectivity, storage or computing resources are limited.
According to Cactus Compute, Whistle processes 16 kHz mono audio clips up to 30 seconds in English, German, French, Spanish, Italian, Dutch and Polish. It detects the spoken language unless the user specifies one. The company says the model can also provide each word’s start and end times, along with a probability score, or return encoder embeddings without generating a transcript.
The model runs in the company’s C++ engine without external dependencies and shares that engine’s components with Needle, Cactus Compute’s existing model. The release describes a five-beam decoding process, keyword biasing for supplied phrases and an option to select decoder depth when loading the model. The encoder still runs all eight of its blocks at every selected depth, according to the technical description.
Cactus Compute reports a 11.1 millisecond time to first token for a 10-second audio clip on an Apple M4 Pro CPU. It also reports 1,319 decoded tokens per second in that test. These figures come from the company’s comparison, which used each model’s official runtime and default settings; they are not a guarantee of speed on other hardware or with different workloads.
Speech Processing on Small Devices
A model that fits in 16.9 MB could make speech recognition more practical on devices with limited storage or intermittent internet access. Cactus Compute lists mobile devices, wearables, robots, smart-home products, vehicles and microcontrollers as intended uses. Local processing can also reduce the need to send audio to a remote service, though the company’s deployment claims do not establish how the model performs across different devices or real-world environments.
The small file size is not the only point of comparison. In its published table, Cactus Compute says Whistle has lower word error rates than Whisper base on LibriSpeech test-clean and test-other, SPGISpeech, Earnings-22 and the FLEURS average. Whisper base performs better on TED-LIUM, AMI and the MLS average. The results indicate dataset-dependent trade-offs, not a single ranking that applies to every speech-recognition task.
The reported speed comparison also needs its test conditions. It covers 10 seconds of audio on an Apple M4 Pro CPU, with Whistle’s C++ engine at five beams, openai-whisper and moonshine-voice using their respective official runtimes and defaults. The numbers are useful as a company-reported comparison under those conditions, but performance on a phone, vehicle processor or microcontroller may differ.
As an affiliate, we earn on qualifying purchases.
How Whistle Fits With Needle
Cactus Compute presents Whistle as an audio model built around the same CPU engine and selected neural-network components as Needle. The company says blocks marked as shared run Needle’s code rather than a separate copy. Whistle adds speech-specific processing, including gated cross-attention that lets its decoder use representations created by the audio encoder.
The reported benchmark compares Whistle with Whisper base and Moonshine tiny v2. The source notes limits in the available comparisons: Moonshine is English-only, and its developers have not published results for some listed datasets. Whisper has no reported figures for SPGISpeech, Earnings-22 or cleaned AMI in the table; its AMI result uses AMI-IHM, a different subset from the AMI result reported for the other systems. Those gaps mean the table is not a complete, directly matched evaluation across every dataset.
Cactus Compute also describes a browser demonstration in which a first-time user downloads the model file and can record a short clip for local transcription. The company says audio in that demonstration stays on the device after download. That is a statement about the demonstration and the product’s intended operation, not an independent audit of all possible integrations.
““Speak, and Whistle transcribes it on your device.””
— Cactus Compute, in its Whistle announcement
portable speech recognition device
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Device Results
The published announcement is the basis for the size, speed and accuracy claims presented here. The supplied material does not include an independent replication of the benchmarks, details on test repetitions or uncertainty ranges, or results across a broad range of target devices. It also does not provide a full numerical word-error-rate table in the report text supplied, so the size of the differences between systems cannot be assessed from those descriptions alone.
It remains unclear how Whistle performs on noisy speech, accents, overlapping speakers and longer or continuous audio beyond the stated 30-second limit. The source does not establish whether all seven languages have similar accuracy, or how results change with decoder depth and keyword biasing. Cactus Compute’s statement that audio stays on-device in its demonstration has not been independently verified in the material provided.
speech transcription app for mobile
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Availability and Wider Evaluations
Cactus Compute’s announcement presents Whistle as released and includes a browser-based demonstration that downloads the model for local use. The source does not specify pricing, a separate distribution channel, license terms or a schedule for updates, so those details remain to be confirmed by the company.
The next useful evidence will be reproducible tests on more hardware, with fuller benchmark figures and results across languages and real-world recording conditions. Until then, developers considering the model can treat the company’s published results as an initial performance report and check them against their own devices, audio and latency requirements.
small footprint speech recognition model
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 released by Cactus Compute. The company says it runs on a CPU through its C++ engine and is packaged as a 16.9 MB file.
Which languages does Whistle support?
Cactus Compute lists English, German, French, Spanish, Italian, Dutch and Polish. The model detects the language unless a user names it, according to the announcement.
Does Whistle process audio online?
The company says processing happens on the device and that audio in its browser demonstration does not leave the device. The demonstration first downloads the model file; the supplied material does not include an independent privacy audit.
How does Whistle compare with Whisper?
Cactus Compute reports that Whistle leads on several listed datasets, while Whisper base leads on TED-LIUM, AMI and the MLS average. Some datasets lack results for one or more models, and the AMI subsets are not the same, limiting direct comparisons.
What are the reported speed figures based on?
The company measured the systems on 10 seconds of audio using an Apple M4 Pro CPU and each model’s stated runtime and default settings. Results may vary on other hardware and under different configurations.
Source: hn
Columbus Day / Indigenous Peoples' Day Picks
long weekend sales
As an affiliate, we earn on qualifying purchases.
