[Models](/models)/Cactus-Compute/whistle

Apache-2.0

# whistle

by Cactus-Compute

A 16.9 MB on-device speech recogniser for 7 languages with word timestamps, 4.31% WER on LibriSpeech clean

129likes

16.9 MB model file (2 to 4 bit)params

[HuggingFace](https://huggingface.co/Cactus-Compute/whistle)[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

Identifiers

Model ID

`Cactus-Compute/whistle`

## Deploy whistle

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

[Read the customer-hosted guide](https://mixpeek.com/docs/customer-hosted/overview) [Talk to us about single-tenant](/demo)

## Overview

Whistle is a speech-to-text model from Cactus Compute, released in September 2026 under Apache-2.0, built for phones, wearables, robots, cars and microcontrollers. The whole model is one 16.9 MB file that runs on Cactus's CPU engine, Needle, with no network.

It transcribes English, German, French, Spanish, Italian, Dutch and Polish, returns every word with its timing, and can also output a speech embedding for matching. On the card's tests it scores 4.31% WER on LibriSpeech test-clean and reaches its first token in 11.1 ms on a 10-second clip.

## Architecture

A log-mel front end and a convolutional stem feed an audio encoder, which a Needle-style decoder reads through gated cross-attention at every layer. Word timestamps come from the decoder's own attention. Weights are quantized to 2 to 4 bits in Cactus's .cact container and run on SIMD CPU kernels.

## Mixpeek SDK Integration

IngestRetrieve

\# Transcribe on the device, then store each clip's transcript as a text object
# with its timing so a Mixpeek retriever can find the moment.
import needle, requests

out = needle.transcribe("clip.wav", word\_timestamps=True)
requests.post(
    "https://api.mixpeek.com/v1/buckets/bkt\_your\_bucket/objects",
    headers={"Authorization": "Bearer API\_KEY", "X-Namespace": "ns\_your\_namespace"},
    json={"blobs": \[{"property": "transcript", "type": "text", "data": out\["text"\]}\],
          "metadata": {"recording": "s3://voice-notes/2026-10-07-0912.wav", "language": out.get("language")}},
)

## Capabilities

*   Transcription of 16 kHz mono audio, up to 30 seconds per pass, in 7 languages
*   Word timestamps with a probability per word
*   Speech embeddings, one row per 80 ms frame, for matching without a transcript
*   Keyword biasing for names, places and product words
*   Runs on CPUs, phones, wearables and microcontrollers, and as WebAssembly

## Use Cases on Mixpeek

Voice commands and dictation on phones, wearables and in cars, with no network

Timestamped transcripts of short clips for search, captions or cutting on a word

Matching spoken phrases by embedding where a transcript is not needed

## Benchmarks

Dataset

Metric

Score

Source

LibriSpeech test-clean

WER (lower is better)

4.31%

Model card: Cactus-Compute/whistle (self-reported, full test split, Whisper normalizers)

LibriSpeech test-other

WER

10.49%

Model card (self-reported)

TED-LIUM

WER

7.61%

Model card (self-reported)

FLEURS, 7 languages

WER

21.4%

Model card (self-reported)

10 s clip, Apple M4 Pro

Time to first token

11.1 ms

Model card (self-reported; Whisper base 73.2 ms, Moonshine tiny v2 22.8 ms)

## Performance

Input Size16 kHz mono audio, up to 30 s per pass

Embedding Dimn/a (text with word timestamps; optional speech embedding per 80 ms frame)

GPU LatencyInput dependent

GPU ThroughputBatch dependent

GPU MemoryModel dependent

16.9 MB against 145.3 MB for Whisper base and 41.9 MB for Moonshine tiny v2, decoding 1,319 tokens a second on a 10 s clip on an M4 Pro (card figures). Longer recordings need to be cut into 30-second pieces. We have not measured it.

## Common Pipeline Companions

[

moondream/parakeet-redux

A larger on-CPU recogniser covering 25 languages, 178 MB

](/model/moondream/parakeet-redux)[

openai/whisper-large-v3-turbo

The server-side choice when size does not matter and you need 100+ languages

](/model/openai/whisper-large-v3-turbo)

## Frequently Asked Questions

### Does Whistle need a GPU or an internet connection?

No. It runs on CPUs, including phones, wearables and microcontrollers, and as a WebAssembly component, with no network calls.

### Which languages does Whistle support?

English, German, French, Spanish, Italian, Dutch and Polish. It detects the language unless you name it.

### How does Whistle compare with Whisper?

On the card's figures it is about a ninth of the size of Whisper base (16.9 MB against 145.3 MB) and reaches the first token about 6 times sooner on a 10-second clip, with 4.31% WER on LibriSpeech clean. The card uses Whisper's published error rates for the comparison.

### How do I search Whistle transcripts with Mixpeek?

Store each transcript as a text object with the recording and its timing in metadata, then search it with a text retriever, as in the example on this page.

## Explore on Mixpeek

[

Browse all models

Search the full model catalog

](/models)[

Feature Extractors

See how Mixpeek runs models as extractors

](/extractors)[

Mixpeek Vector Store

Store & search embeddings at scale

](/mvs)[

Pricing

Usage-based pricing for pipelines

](/pricing)[

All Curated Lists

Compare models, APIs & infrastructure

](/curated-lists)

### Specification

OrganizationCactus-Compute

Retriever\-

Parameters16.9 MB model file (2 to 4 bit)

LicenseApache-2.0

Downloads/moN/A

Likes129

[

### Research Paper

Cactus-Compute/whistle model card

arxiv.org](https://huggingface.co/Cactus-Compute/whistle)

### Build a pipeline with whistle

Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)