Vosk

Vosk is an open-source speech recognition toolkit for applications that need offline transcription. It works on Android, iOS, Raspberry Pi, and other listed systems, and supports more than 20 languages and dialects. A streaming API handles ongoing audio recognition, while vocabulary reconfiguration lets developers adjust supported terms; most small models allow vocabulary changes while running. The toolkit has speaker identification, timestamps, and TXT and SRT export. Small models are typically around 50 MB and use about 300 MB of runtime memory; large server models can require up to 16 GB. Developers can use bindings for Python, Java, C#, and JavaScript, among other languages. Vosk Server has WebSocket, gRPC, MQTT, and WebRTC implementations, and listed integrations include Asterisk, FreeSWITCH, Jigasi, and UniMRCP. The server can be installed with Docker and use a locally mounted model folder. The free plan costs nothing. The installation page lists ARMv6 and Windows ARM64 as unsupported, and iOS builds are available on request.

Vosk provides offline speech recognition with streaming, speaker identification, timestamps, and multiple developer interfaces. Its model memory needs and stated platform limits are key considerations when choosing a deployment.

Vosk suits developers building speech recognition into applications that need offline operation, streaming, or configurable vocabulary. Its listed bindings, server protocols, and mobile and lightweight-device support offer several deployment options.

Vosk is an open-source toolkit for developers embedding speech recognition in their own software. It is a strong fit for teams that need local, continuous recognition across devices and can manage model deployment themselves. The trade-off is that hardware requirements and platform caveats demand more engineering judgment than a hosted transcription service.

Overview

Vosk is a library rather than a finished transcription workspace. It recognizes speech offline in more than 20 languages and dialects, including on Raspberry Pi, Android, and iOS. That local operation suits products that cannot depend on a network connection or should keep audio processing on the device. The FAQ positions it as a practical, plug-and-play application library; teams pursuing speech-recognition research should look to Kaldi instead.

Model selection directly affects deployment. Small models are typically about 50 MB and require roughly 300 MB of runtime memory, while large server models can need up to 16 GB. Vosk can therefore serve lightweight hardware and substantial servers, but there is no single model profile that fits every target.

Key features

Offline streaming recognition

The streaming API supports ongoing speech input without treating internet access as a prerequisite. Speaker identification, speaker labels, and timestamps add structure for applications handling conversations or recorded sessions. TXT and SRT export cover straightforward text and subtitle workflows, although Vosk remains primarily a component for developers rather than an end-user editing suite.

Language and vocabulary control

Support for more than 20 languages and dialects broadens its use beyond a single-language product. Vocabulary reconfiguration is another practical advantage: the models page says most small models support dynamic vocabulary changes, allowing an application to adapt recognition to its terminology instead of relying only on a fixed lexicon.

Interfaces and server protocols

Bindings include Python, Java, C#, and JavaScript, among other languages, so teams can use Vosk from common application stacks. Vosk Server implementations cover WebSocket, gRPC, MQTT, and WebRTC. Docker installation with a locally mounted model folder makes repeatable, self-hosted deployment more manageable. Listed integrations include Asterisk, FreeSWITCH, Jigasi, and UniMRCP, which is useful for telephony and real-time communications projects.

Pricing

There is no subscription charge: Vosk is open source and the API repository uses the Apache-2.0 license. That makes it suitable for teams able to fund their own integration, hosting, model management, and support work. It does not provide the convenience of a hosted service with a metered quota or packaged workflow; the engineering and operating responsibility stays with the adopter.

Platforms

The directory lists Android, iOS, Linux, macOS, Windows, self-hosted deployment, and API access. The Python package installation page lists Linux x86_64, Linux arm64, Raspberry Pi 3 and 4, macOS x86 and M1, and Windows x86 and 64. ARMv6 and Windows ARM64 are not supported, while iOS builds are available on request. Those boundaries matter when a product must run on a specific embedded board or Windows architecture.

Support is routed through GitHub issues, email, Telegram, and an invitation-only WeChat group. This is a community and project support model rather than a guaranteed commercial help desk.

Who it's for

Choose Vosk for an application team building offline or streaming speech recognition into a mobile app, embedded device, communications system, or self-hosted server. It is especially compelling when local processing, configurable vocabulary, speaker labeling, or control over model placement matters. It is less suitable for buyers seeking a turnkey transcription workspace, managed collaboration features, or an interface designed primarily for nontechnical users.

Pros and cons

• Pros: offline operation keeps recognition viable without connectivity; streaming, timestamps, and speaker identification support real-time and structured audio workflows; bindings and server protocols cover varied stacks; small models can fit lightweight devices; Apache-2.0 licensing removes a software subscription cost.
• Cons: large server models may require up to 16 GB of memory; ARMv6 and Windows ARM64 are unsupported; iOS builds require a request; deployment, model selection, and ongoing operation are left to the adopting team.

Alternatives

For a hosted transcription product with a user-facing workflow, consider Happy Scribe. MacWhisper is the more focused choice for macOS and iOS users wanting transcription and subtitle export, with a free tier and a €64 a month Pro plan billed as Pay Once. Notta suits teams wanting a broader cross-platform service with a free plan capped at one seat and 120 transcription minutes per month.

tl;dv is aimed at meeting recording and transcription, including unlimited recordings and transcriptions on its free plan. Transkriptor fits individuals seeking a hosted monthly allowance, with a Lite plan at $9.99 a month for 300 minutes. Otter.ai is another hosted option with workspace and monthly transcription limits on its Basic plan. Google Voice Access is a free Android-only alternative. Zoom Scheduler is for booking pages and calendar connections, not a direct speech-recognition substitute.

Verdict

Vosk is the right choice for developers who need free, offline, streaming speech recognition and want control over models and deployment. Its combination of device reach, language support, vocabulary changes, and developer interfaces is the reason to choose it. Look elsewhere if you want a polished hosted transcription service or cannot accommodate its model memory requirements and platform restrictions.