Whistle

Whistle

Transcribes speech in seven languages with word timestamps and speech embeddings, in a model only 16.9 MB in size. Runs from a short Python script, for developers adding speech to apps.

Made by Cactus-Compute on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Whistle is a tiny speech model from Cactus Compute. It turns audio into text, gives start and end times for words, and produces speech embeddings, all from a model of about 16.9 MB.

It covers English, German, French, Spanish, Italian, Dutch and Polish. On first use, a Python package fetches the engine and the model weights and keeps a local copy.

Who it suits: Developers who want speech features inside an app or on a small device, and people who care about a small download and local processing.

Industry | Who uses it | How they use it | Example
Software | App developers | Call one function to transcribe audio | Turn a recorded clip into text
Devices | Teams building for small hardware | Run speech recognition with a small footprint | Add voice input to a compact gadget
Media | Editors and archivists | Get word timings from recordings | Mark where each word starts and ends
Research | Speech researchers | Use speech embeddings from a small model | Compare clips by their embeddings

Skip it if you need a language outside the seven listed, since the card does not cover others.

Where to use it:

• Python programs on Windows, macOS or Linux
• A browser demo on the Whistle website
• Microphone capture, after installing the extra package option
• Platforms listed in the Cactus runtime repository

Use case | Who | What they get back
Transcription | Developers | The spoken text of a 16 kHz WAV file
Word timestamps | Subtitle and editing teams | Start and end times for each word
Speech embeddings | Researchers and search builders | Vectors that describe the audio

How to set it up:

• Python and pip
• A 16 kHz WAV file to transcribe
• Internet access the first time, so the package can fetch and cache the engine and weights

1. Install the cactus-compute package.
2. Put a 16 kHz WAV file where your script can reach it.
3. Run a short script that calls the transcribe function.

pip install -U cactus-compute

How to use it: Import cactus and call transcribe with the path of a WAV file, then print the text from the result. For other sample rates or microphone input, install the package with its mic option.

import cactus

print(cactus.transcribe("clip.wav")["text"])

Tool | What it does
cactus.transcribe | Transcribes an audio file and returns a result that holds the text
whistle.cact | The model file this repository provides
checkpoints/whistle.safetensors | The checkpoint file this repository provides

Example: You save a short recording as clip.wav at 16 kHz and run the script above. It prints the words spoken in the clip as plain text.

For newcomers and experts: If you are new to speech recognition, install the package and run the three-line script on one short clip. Start with a clear recording in one of the seven languages.

If you know the field, the model is a compact checkpoint of about 16.9 MB that handles transcription, word timestamps and speech embeddings. The repository ships it as a cact file and a safetensors checkpoint, and the card links guides for the Python API, speech use and fine-tuning.

16 kHz WAV file or microphone
cactus.transcribe call
Engine and weights cached on first use
Text, word times or embeddings

Good to know:

• The default input is a 16 kHz WAV file
• Runtime binaries and supported platforms are kept in the separate Cactus repository
• Hugging Face lists 5 quantized and 13 fine-tuned versions made by other people
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the audio files or microphone input you give it, plus a first-use download of the engine and weights
• Not for: other sample rates without the extra mic option, or languages outside the seven listed

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: cactus-compute
• Downloads in the last 30 days: 94,489
• Likes: 570
• License: apache-2.0
• Languages: en, de, fr, es, it, nl, pl
• Quantized versions: 5
• Fine-tunes: 13
• Tags: cactus-compute, whistle, speech-recognition, speech-to-text, on-device, edge, quantization, webassembly, automatic-speech-recognition

Links

• Model: https://huggingface.co/Cactus-Compute/whistle
• Files: https://huggingface.co/Cactus-Compute/whistle/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Cactus-Compute/whistle

Then follow the usage in its model card. Nothing in this kit runs by itself.