Whistle

Use it to transcribe speech, align words with timestamps, or create speech embeddings on devices such as mobiles, wearables, robots, and microcontrollers.

Availability

The model file and source are available on Hugging Face and GitHub.

Pricing model

Not established

Deployment

On-device; no GPU or dependencies required for the model.

Privacy

Transcription runs on the device.

Licence

Not established

• Fine-tuning adapts Whistle to speakers and vocabulary or teaches it a new language. ^[1]

  needle whistle finetune adapts it to your speakers and vocabulary, or teaches it a new language, on your own machine.

• Fine-tuning runs on the user's own machine. ^[1]
• The build command merges the adapter and writes a .cact file that carries the languages and tokens. ^[1]

  build merges the adapter and writes a .cact that carries your languages and tokens.

• Streaming, keywords and word timestamps work as they do with the shipped model. ^[1]

  Streaming, keywords and word timestamps work as they do with the shipped model.

• Brazilian Portuguese fine-tuned with about ten hours of FLEURS data reached 30.5% word error rate on the FLEURS test set. ^[1]

  Brazilian Portuguese from about ten hours of FLEURS, at --lr 1e-2 with the other defaults, went from 106.2% to 30.5% word error rate on FLEURS test and was detected as Portuguese on every test clip.

• In the browser sandbox, audio never leaves the device. ^[2]

  The first press downloads the 16.9 MB model, and audio never leaves your device.

• Whistle transcribes 16 kHz mono audio in English, German, French, Spanish, Italian, Dutch, and Polish. ^[3]

  - **Transcription**: 16 kHz mono audio of any length, one pass up to 30 seconds and windowed beyond it, in English, German, French, Spanish, Italian, Dutch and Polish.

• Whistle provides word start and end times and a probability for every word. ^[3]

  - **Word timestamps**: every word with its start, end and probability, aligned from the decoder's own attention, so an app can highlight, seek or cut on a word.

• Whistle provides speech embeddings with one row per 80 ms frame. ^[3]

  - **Speech embedding**: the encoder output, one row per 80 ms frame, for matching and retrieval without decoding a transcript at all.

• The repository contains the downloadable whistle.cact model file. ^[3]
• Whistle runs on-device on a CPU without a GPU. ^[3]

  The whole model is a single 16.9 MB file, and it runs on the same CPU engine as [Needle](https://huggingface.co/Cactus-Compute/needle3), from the same container and the same quantisation, with no dependencies and no GPU.

• The complete model is a single 16.9 MB file. ^[3]

  The whole model is a single 16.9 MB file, and it runs on the same CPU engine as [Needle](https://huggingface.co/Cactus-Compute/needle3), from the same container and the same quantisation, with no dependencies and no GPU.

• Whistle can now be tuned for a speaker or new language Daily brief 2026-10-18
• Backburner leads 25 AI developments in week 40, 2026 Weekly brief 2026-W40
• Google introduces Gemini 4 Argon reasoning model Daily brief 2026-10-02