Whistle: On-Device Speech-to-Text in Just 16.9 MB

X postMario ZechnerTool

Mario Zechner highlights Whistle by Cactus, an open speech-to-text model requiring only 16.9 MB of storage. It runs CPU-only without external dependencies and supports transcription in seven languages.

Key points

  1. Whistle is a compact 16.9 MB speech recognition model tailored for mobile devices, wearables, robotics, and microcontrollers.
  2. Transcribes up to 30 seconds of audio in English, German, French, Spanish, Italian, Dutch, and Polish, providing word timestamps and speech embeddings.
  3. Runs on the same C++ engine as the Needle 3 foundation model, enabling direct audio-to-tool-call execution in a single binary.
  4. Delivers a time-to-first-token of 11.1 ms on an Apple M4 Pro CPU, decoding at 1,319 tokens per second.
  5. User feedback indicates potential recognition issues with certain accents, though explicit language selection improves accuracy.

Why it matters

Local on-device voice processing has traditionally been constrained by model size and latency. A functional 16.9 MB speech model running purely on CPU enables private, offline voice interfaces on low-power hardware without relying on cloud services.

Analysis

Cactus benchmarks Whistle against Whisper base (145.3 MB) and Moonshine tiny v2 (41.9 MB), showcasing lower latency and smaller footprints. The architectural synergy with Needle allows pipelines to convert spoken input straight into structured tool calls within one binary. However, community tests note that automatic language detection can struggle with non-standard accents.

Original post

Mario Zechner

@badlogicgames · 8 October 2026

amazing t.co/CU8Y5jIRcT

526 likes29 replies592 bookmarks24.3k views

View on X
Selected replies (5)
  • @calebfahlgren @badlogicgames t.co/eNGaVTwSaY
  • @sridca @badlogicgames not amazed, because it sucks at non-Western English accents
  • @LetMeKeyboard @badlogicgames I was testing its recognition of my non-ideal accent, and then just countries in general, and this is what i got for "wtf is Montenegro". Amazing indeed. t.co/cylCAl66Fr
  • @gkurttech @badlogicgames At first I thought it was bad. But then I selected English explicitly and it got very good. Maybe the language detection is not good.
  • @i_m_re @badlogicgames I just tried some Dutch on my phone and it was blazingly fast and 100% accurate. Also for very short sentences. Amazing work

Mentioned links

Links and tools from this entry

Summary generated by AI (Gemini 3.8 Flash, 9 October 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Article:Jakub Mroz, Henry Ndubuaku

    Whistle: Open-Source Speech-to-Text in 16.9 MB

    Cactus Compute has released Whistle, a lightweight speech recognition model packaged into a single 16.9 MB file. Running on CPU without external dependencies, it shares its C++ engine with Needle to convert audio directly into tool calls.

    AI & LLMs· Announcement

  • Repository:google-ai-edge/LiteRT-LM

    LiteRT-LM: Google's Open-Source Inference Framework for On-Device LLMs

    LiteRT-LM is an open-source inference orchestration layer from Google designed to run large language models directly on edge devices. It leverages GPU and NPU acceleration and powers on-device generative AI features in Chrome, Pixel Watch, and Chromebook Plus.

    6.4kstarsC++

    AI & LLMs· Tool

  • Repository:huggingface/speech-to-speech

    Hugging Face Speech to Speech: Modular Open-Source Voice Agent Pipeline

    Hugging Face has released speech-to-speech, a modular open-source framework designed to build voice agents with open-source models. The pipeline spans VAD, STT, LLM, and TTS, exposing the core OpenAI Realtime GA event set over WebSocket and WebRTC.

    13kstarsPython

    AI & LLMs· Tool

  • Repository:jamiepine/voicebox

    Voicebox: Open-Source Local AI Voice Studio and Dictation Tool

    Voicebox is a local-first, open-source voice studio combining text-to-speech, voice cloning, and system-wide dictation. It serves as a self-hosted alternative to proprietary cloud solutions such as ElevenLabs and WisprFlow.

    55.6kstarsTypeScript

    AI & LLMs· Tool

  • X post:MIKE

    Open-Source Local Alternative to ElevenLabs for Voice Cloning and TTS

    An X post highlights an open-source project by a solo developer that runs text-to-speech and voice cloning entirely on local hardware. The repository has accumulated 19.4k GitHub stars, presenting itself as a free self-hosted alternative to ElevenLabs.

    21.1kbookmarks1.1mviews

    AI & LLMs· Tool

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.