VOSK

An offline, open-source STT toolkit from Alpha Cephei, built on the Kaldi ASR framework and DNN acoustic models. Its niche is the opposite of the accuracy leaderboard: tiny footprint, permissive license, and runs anywhere. Grounded here in the primary vosk-api repo (~15k GitHub stars; latest release v0.3.50, April 2024 — a mature, slow-moving project).

What the repo documents

• ~50 MB models, with larger server-grade variants available — the compactness is the headline.
• 20+ languages and dialects (English, German, French, Spanish, Chinese, Russian, …).
• Streaming, zero-latency recognition with a reconfigurable vocabulary/grammar, plus speaker identification — so it does more than transcribe (a capability axis beyond WER).
• Runs “from a Raspberry Pi or Android phone to big clusters” — Android, iOS, Pi, and server.
• Bindings for many languages: Python, Java, Node.js, C#, C++, Rust, Go, and more.
• Apache-2.0 — fully permissive, not a research/non-commercial license.

Where it sits

VOSK is the small/embedded, permissively-licensed pole of the open STT field. It does not compete with the SALM-style accuracy leaders (canary-qwen) or even Whisper on WER — it competes on size, speed, offline operation, and running without a GPU. That’s exactly why an app like vocalinux offers it as a backend: on a low-power or GPU-less Linux box, VOSK dictates where a multi-gigabyte Whisper model wouldn’t be practical. Two threads land on it. It fills the corner the wiki’s WER-focused pages (speech-to-text, open-source-stt-models) leave open — models chosen for constraint, not benchmark rank. And on the spoke’s license axis, VOSK sits at the clean end: where the TTS open leader fish-audio-s2-pro hits a research-license ceiling, VOSK is Apache-2.0 — the “best-sounding option is often the least legally safe” tension doesn’t bite in the small-STT corner, where the constrained option is also the freely-shippable one.

Caveats

The README emphasizes DNN-based recognition and doesn’t spell out the Kaldi lineage in the fetched text; the Kaldi foundation is established public record for the project. WER/accuracy figures aren’t given — VOSK’s pitch is footprint and licensing, not topping the Open ASR Leaderboard.

Related

speech-to-text · vocalinux · whisper · open-source-stt-models · fish-audio-s2-pro · speech-audio-ai