What is Vosk speech recognition?

Vosk is an open-source speech recognition toolkit made by Alpha Cephei. It runs offline, including on lightweight devices such as the Raspberry Pi, Android and iOS ¹. It supports more than 20 languages and dialects ², and its per-language models are about 50 MB ³. It is best suited to small devices and fixed command sets. For long, noisy dictation, test it against a neural model before you choose it.

Vosk is one of the older toolkits covered in our guide to open-source speech recognition, which compares it with the newer neural models.

Vosk is an offline toolkit with models of about 50 MB per language

The Vosk project says it works offline even on lightweight devices, including the Raspberry Pi, Android and iOS ¹. It supports more than 20 languages and dialects, among them English, Indian English, German and French ². Each per-language model is about 50 MB, and the project also offers much bigger server models ³.

The Vosk API is released under the Apache License, Version 2.0 ⁴. Apache 2.0 is permissive, but check the license of each language model before you ship, since models for other languages can carry different terms.

Vosk's word error rates are its own measurements

Word error rate (WER) is the share of words a system gets wrong, so lower is better. The Vosk project's model table gives its small English model, vosk-model-small-en-us-0.15, a 9.85 percent WER on LibriSpeech test-clean, a read-audiobook test set ⁵. Its large English model, vosk-model-en-us-0.22, scored 5.69 percent on the same set ⁶.

These are the project's own figures on one test set. They are not on the Open ASR Leaderboard, so you cannot rank them against Parakeet, Canary or Whisper from these numbers. The project's accuracy guide says that modern systems, Vosk included, can be very accurate at one time and poor at another, because audio quality, vocabulary and accent all matter ⁷. Whether Vosk is accurate enough for open-ended dictation of long, accented or noisy speech is a question only your own tests can answer.

Memory decides which Vosk model you can run

Memory is the clearest guide to the model you need. The project says a typical small model is about 50 MB on disk and needs about 300 MB of memory at runtime ⁸. Its big models "require up to 16Gb in memory since they apply advanced AI algorithms" ⁹.

So a phone, a Raspberry Pi or a small laptop points to a small model, and a server points to a big one. Check the free memory on your device before you choose.

Vosk fits fixed commands and small devices

Vosk is most useful where a neural model is too large for the hardware, or where the set of words is small and known in advance, such as a list of voice commands. It is a practical choice for a board that controls equipment, or an app that must work with no network.

For open-ended dictation of long, accented or noisy speech, I did not find a head-to-head test of Vosk and the neural models on the same audio. The choice for that job should rest on your own test, not on a claim from either side.

What to do: test Vosk on your device and your recordings

1. Check the device's memory against the roughly 300 MB a small model needs at runtime ⁸.
2. Record 20 to 30 clips that match your real use, and transcribe them by hand.
3. Compare WER against a neural model on the same clips. If the neural model is too large for the device, compare against the best model that fits.
4. Keep Vosk only if its error rate on your clips is acceptable for the job.

If you want the wider picture, read the best alternatives to Whisper and what NVIDIA Parakeet is. For the on-device side of this choice, see the guide to on-device speech recognition. The speech recognition overview places all of these in context.