SSyncDev

Whistle

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Whistle is Cactus Compute's 16.9 MB on-device speech recognition model for CPUs, covering seven European languages with word timestamps and keyword biasing.

Key specifications

Specification score
9.7
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Audio input

Produces

  • Text output

Capabilities

AudioOpen weights

What is Whistle?

16.9 MB. Whistle, released by Cactus Compute on 2 October 2026, is a speech recogniser small enough to ship inside a mobile app without anyone noticing the download. It runs on the CPU with no dependencies and transcribes English, German, French, Spanish, Italian, Dutch and Polish.

For scale: Whisper base is about 145 MB.

Speed on a laptop CPU

On an Apple M4 Pro, transcribing a 10-second clip, Cactus measured:

Model Size First token Decode speed
Whistle 16.9 MB 11.1 ms 1,319 tokens/s
Moonshine tiny v2 41.9 MB 22.8 ms 262 tokens/s
Whisper base 145.3 MB 73.2 ms 266 tokens/s

Time to first token grows with clip length: about 5.9 ms at 5 seconds, 36.3 ms at 30. These are the vendor's measurements on its own hardware.

Accuracy is mixed, and Cactus says so. It reports Whistle ahead on LibriSpeech, SPGISpeech, Earnings-22 and the FLEURS average, while Whisper base still wins on TED-LIUM, AMI meeting audio and the MLS average. Exact word error rates appear only in charts, and no parameter count is published.

More than a transcript

  • Word timestamps with a start, end and probability for every word.
  • Speech embeddings, one row per 80 ms frame, without decoding a transcript.
  • Keyword biasing to push product names or jargon up the likelihood list.
  • Automatic language detection, or a fixed language if you know it.

Input is 16 kHz mono; one pass covers 30 seconds and longer audio is windowed. Silent clips return an empty transcript without decoding, which saves battery on always-listening features.

Where it runs

The Cactus engine ships prebuilt for 17 targets, including iOS, Android, macOS, Linux, Windows on ARM, browsers and WASI. In Python it is pip install cactus-compute, then a single transcribe call. On the command line, cactus download whistle fetches the model.

The weights carry an Apache-2.0 licence on their repository.

Picking it, or not

Whistle makes sense when audio must stay on the device, when you need a result in milliseconds, or when app size matters. It makes less sense for meeting transcription with crosstalk, where its own numbers trail Whisper base, or for languages outside its seven.

Test it with your microphones, not studio audio. Record 30 clips in the noise your users live with, transcribe them on your slowest supported phone, and compare word error rate and battery draw with your current recogniser. Then try keyword biasing on the five terms you most often see misspelled.

Larger alternatives sit under speech recognition; Whisper Large v3 Turbo is the usual server-side yardstick.

How to evaluate Whistle

Compare the model on the work you need to complete.

Tasks and inputs

Whistle is listed for Automatic speech recognition, Speech → Text. It accepts audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

English, German, French, Spanish, Italian, Dutch and Polish, with automatic language detection.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

salesforce

SOV-Transcribe-1

SOV-Transcribe-1 is Salesforce's open-weight speech-augmented model for non-streaming transcription, with keyword biasing and style instructions; Apache-2.0.

— context

ElevenLabs

Scribe v3 Live

Scribe v3 Live is ElevenLabs' streaming transcription model with turn prediction, mid-sentence language switching and a 10,000-term keyterm library.

— context

NVIDIA

Nemotron 3.5 Transcribe 30B A4B

Nemotron 3.5 Transcribe 30B A4B is NVIDIA's open offline ASR model: 40 languages, hour-long files, word timestamps, diarisation and translation.

65.54K context

IBM

Granite Speech 5.0 2B

Granite Speech 5.0 2B is IBM's open speech-language model for multilingual ASR and translation with keyword biasing, runnable from Ollama to MLX.

— context

Mistral AI

Voxtral Mini Transcribe 3

Voxtral Mini Transcribe 3 is Mistral's 4B batch transcription model with speaker labels, segment timestamps and custom vocabulary, replacing Mini 2.

65.54K context

Mistral AI

Voxtral Mini Transcribe Realtime 3

Voxtral Mini Transcribe Realtime 3 is Mistral's streaming speech-to-text model for voice agents: under 500 ms, tunable delay, built-in turn detection.

— context