Overview
Vosk is an offline, open-source speech recognition toolkit built around small models and wide portability. Its models are about 50 MB yet still do continuous large-vocabulary transcription, with a streaming API that returns results with zero latency rather than waiting for an utterance to end. It also supports reconfigurable vocabulary — narrowing the recognised word set to a domain — and speaker identification.
Coverage runs to more than 20 languages and dialects, including English and Indian English, German, French, Spanish, Portuguese, Chinese, Russian, Turkish, Vietnamese, Italian, Dutch, Catalan, Arabic, Greek, Farsi, Filipino, Ukrainian, Kazakh, Swedish, Japanese, Esperanto, Hindi, Czech and Polish.
The deployment story is the reason it shows up in embedded and on-device work: bindings exist for Python, Java, Node.js, C#, C++, Rust, Go and others, and the same toolkit scales from a Raspberry Pi or an Android phone up to a server cluster. Typical uses are chatbots, smart home appliances and virtual assistants, plus subtitle generation for video and transcription of lectures and interviews.
The project is Apache-2.0 licensed and maintained by Alpha Cephei; installation instructions, model downloads and examples live on the Vosk website rather than in the repository README.
What it does
- Fully offline — no network call and no data leaving the device
- ~50 MB models that still do continuous large-vocabulary transcription
- Streaming API with zero-latency partial results
- 20+ languages and dialects
- Reconfigurable vocabulary to bias recognition toward a domain
- Speaker identification
- Bindings for Python, Java, Node.js, C#, C++, Rust, Go and others
- Runs on Raspberry Pi and Android through to server clusters
Getting started
Installation, model downloads, examples and full documentation live on the Vosk website; the repository README points there rather than duplicating it.
Read the install and model docs
Pick a language model and the binding for your platform from the official site — models are downloaded separately from the API package.
https://alphacephei.com/voskGet the source
The repository carries the API implementations, language bindings and examples.
git clone https://github.com/alphacep/vosk-api.gitCommands and code are distilled from the project's own documentation — always check the official repo for the latest.
When to use it
- On-device transcription for phones, Raspberry Pi and other hardware with no reliable network
- Voice control for smart-home appliances and virtual assistants where latency and privacy both matter
- Generating subtitles for video, or transcribing lectures and interviews, without sending audio to a provider
- Domain-specific recognition where a reconfigured, narrower vocabulary beats a general model
How Vosk compares
Vosk alongside other open-source audio, music & voice tools AI/TLDR tracks, ranked by GitHub stars.
| Tool | Stars | What it does |
|---|---|---|
| Whisper | ★ 109k | OpenAI's speech recognition model that transcribes and translates audio across many languages. |
| GPT-SoVITS | ★ 61.7k | An open-source WebUI that clones a voice from a short audio sample and turns text into speech, with zero-shot and few-shot fine-tuning. |
| VibeVoice | ★ 54.2k | Microsoft's text-to-speech model for generating long, expressive multi-speaker audio like podcasts. |
| whisper.cpp | ★ 53.6k | A dependency-free C/C++ port of Whisper built on ggml, running speech recognition on CPU, Metal, CUDA, Vulkan and NPUs from phones to servers. |
| Voicebox | ★ 53k | Local-first voice studio that clones a voice from a short sample, generates speech across seven TTS engines and 23 languages, handles system-wide dictation, and speaks for agents over MCP. |
| Coqui TTS | ★ 46k | A library of text-to-speech models including the multilingual XTTS voice-cloning model. |
| ChatTTS | ★ 39.8k | ChatTTS is an open-source text-to-speech model tuned for dialogue, with multi-speaker support and fine-grained control over laughter, pauses, and prosody. |
| Vosk | ★ 15.1k | Offline speech recognition in 20+ languages from ~50 MB models, with streaming output and bindings for most languages |