Vosk
alphacephei.com
- Free plan
- Windows
- macOS
- Linux
- iPhone and iPad
- Android
- API
- Self-hosted
What Vosk is
Vosk is a free, open-source speech recognition toolkit for applications that need offline transcription. It supports more than 20 languages and dialects and can run on lightweight devices such as Raspberry Pi, Android, and iOS. A streaming API is available, and developers can reconfigure vocabulary; most small models support dynamic vocabulary changes. Small models are typically around 50 MB and need about 300 MB of runtime memory, while large server models can require up to 16 GB. The project provides bindings for Python, Java, C#, JavaScript, and other languages. Vosk Server has implementations for WebSocket, gRPC, MQTT, and WebRTC, with listed integrations for Asterisk, FreeSWITCH, Jigasi, and UniMRCP. The server can be installed with Docker and use a locally mounted model folder. Listed systems include Linux x86_64 and arm64, Raspberry Pi 3/4, macOS x86 and M1, and Windows x86 and 64 for the Python package. ARMv6 and Windows ARM64 are unsupported; iOS builds are available on request. The API repository identifies its license as Apache-2.0.
Who it suits
Vosk suits developers building practical speech recognition into applications, especially when offline operation matters. Its FAQ suggests Kaldi for research use.
Good to know, and what to watch
- Runs offline on lightweight devices.
- Supports more than 20 languages and dialects.
- Includes a streaming API.
- Offers bindings for several programming languages.
- Apache-2.0 licensed API repository.
- Large server models can require up to 16 GB.
- ARMv6 and Windows ARM64 are unsupported.
- iOS builds are available on request.
Plans
Every plan and priceVosk Open Source
Open source
Getting started
- Open the Vosk website
- Choose a supported platform or API
- Install the package for your operating system
- Select and mount a Vosk model locally for server use
- Connect through a language binding or server protocol
- Configure vocabulary, speaker labels, timestamps or export formats
Limits to check
The free Vosk Open Source plan is free and covers offline speech recognition on mobile and server across 10 platforms. Model resource needs vary from roughly 300 MB of runtime memory for typical small models to up to 16 GB for large server models.
The Club's review
Vosk is an open-source toolkit for developers embedding speech recognition in their own software. It suits teams that need local, streaming recognition on anything from a server to a Raspberry Pi. The trade-off is practical rather than cosmetic: model memory requirements and platform gaps must match your deployment.
Overview
Vosk is a library, not a finished transcription workspace. It recognizes speech offline and supports more than 20 languages and dialects, which is valuable for applications that cannot depend on a network connection or should keep audio processing local. The FAQ positions it as a practical, plug-and-play library for applications; teams doing speech-recognition research should look at Kaldi instead.
Model selection directly affects hardware. Typical small models are about 50 MB and need around 300 MB of runtime memory, while large server models can require up to 16 GB. That range makes both embedded and server deployments possible, but it also means accuracy and capacity decisions need to be made alongside infrastructure planning.
Key features
Offline and streaming recognition
Vosk combines offline processing with a streaming API, allowing continuous recognition without making connectivity a prerequisite. That is a strong fit for field devices, privacy-sensitive applications, and installations with unreliable links. Speaker identification, speaker labels, and timestamps add structure to results, while TXT and SRT export cover straightforward text and subtitle workflows.
Vocabulary and language coverage
Vocabulary reconfiguration gives application developers a way to adapt recognition to their terminology. Most small models support dynamic vocabulary changes, useful for focused domains where names or commands matter more than a generic word list. Support for more than 20 languages and dialects broadens deployment options, though the model still has to fit the target device.
Developer and server interfaces
Bindings include Python, Java, C#, and JavaScript, among other languages, so Vosk can sit inside a range of application stacks. Vosk Server offers WebSocket, gRPC, MQTT, and WebRTC implementations for teams that want a service boundary rather than an in-process library. Docker installation with a locally mounted model folder supports repeatable self-hosted deployments. Integrations listed for Asterisk, FreeSWITCH, Jigasi, and UniMRCP are particularly relevant to telephony and communications systems.
Pricing
| Plan | Price | Includes |
|---|---|---|
| Vosk Open Source | Free | Offline speech recognition, mobile and server, 10 platforms |
There is no subscription, usage quota, seat charge, or hosted-service commitment in this plan. The Apache-2.0 license identified by the Vosk API repository permits a broad range of development and deployment use, making it suitable when your team can operate the software and models itself. The cost advantage comes with engineering responsibility: you select models, provide runtime capacity, and maintain the integration rather than buying a managed transcription service.
Platforms
Vosk supports Android, iOS, Linux, macOS, Windows, APIs, and self-hosted deployments. 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. iOS builds are available on request, so iOS work may require an extra delivery step. Small models make lightweight hardware realistic, while large server models can demand up to 16 GB of memory.
Who it's for
Choose Vosk when you are building speech input, transcription, subtitles, or voice-driven workflows into an application and need offline operation, streaming, or local control. It is well suited to embedded products, mobile apps, on-premises services, and communications integrations. It is less suitable if you want a polished, user-facing transcription product with managed hosting and minimal maintenance, or if your target system is ARMv6 or Windows ARM64. Research teams should favor Kaldi for research-oriented work.
Pros and cons
Pros
- Offline operation: recognition can continue on local and lightweight devices without a network dependency.
- Flexible deployment: bindings, server protocols, Docker support, and multiple model sizes cover embedded and server architectures.
- Useful recognition controls: streaming, speaker identification, timestamps, speaker labels, and vocabulary reconfiguration support application-specific workflows.
- No licence fee: the open-source plan costs nothing and uses an Apache-2.0 API repository licence.
Cons
- Memory varies sharply: a large server model may need up to 16 GB, so the free software does not guarantee a low-resource deployment.
- Architecture exclusions: ARMv6 and Windows ARM64 are unsupported, and iOS builds are available on request.
- Operational ownership: self-hosting leaves model selection, capacity, integration, and ongoing support with your team.
Alternatives
For a managed API with paid usage, a free plan, and a free trial across API, macOS, web, and Windows, choose Gladia instead. Superwhisper is the better fit for a ready-to-use desktop and mobile voice-to-text experience with a freemium plan and 100+ languages. Notta suits users who want a hosted workspace with a free tier capped at one seat and 120 transcription minutes per month. Pick AssemblyAI for a paid API with free audio credits and defined streaming and concurrent-transcription limits. CMU Sphinx is another free, BSD-like toolkit with commercial distribution allowed. Kaldi is the stronger alternative for speech-recognition research and also uses an Apache License 2.0. Picovoice Leopard is a paid option with a seven-day, no-card trial. Talon Voice is aimed at voice-control users and offers freemium access.
Verdict
Vosk is a strong choice for developers who need free, offline and streaming speech recognition that can run across mobile, embedded, and self-hosted systems. Its main reason to choose it is deployment control, including vocabulary and protocol flexibility. Look elsewhere if you need a managed transcription product, guaranteed support for an excluded architecture, or a workflow that avoids model and infrastructure decisions.
The verdict
Vosk provides offline recognition, streaming, and vocabulary controls across a range of devices and developer interfaces. Model memory needs vary considerably, and some system architectures are unsupported.
Questions people ask
Is Vosk free?
Yes. The Vosk Open Source plan costs nothing, and the project is open source.
Can Vosk work without internet access?
Yes. It is designed for offline recognition, including on Raspberry Pi, Android and iOS.
Which platforms are supported?
Listed platforms include Android, iOS, Linux, macOS, Windows, APIs and self-hosted deployments. The Python package lists Linux x86_64 and arm64, Raspberry Pi 3/4, macOS x86 and M1, and Windows x86 and 64.
What output and recognition features are available?
Vosk supports speaker identification and labels, timestamps, streaming recognition, and TXT or SRT exports.
Can developers change the vocabulary?
Yes. Vocabulary can be reconfigured, and most small models support dynamic vocabulary changes.
What license does Vosk use?
The Vosk API GitHub repository identifies its license as Apache-2.0.
What the maker publishes
- API access
- Yes source
- Deployment
- The server can be installed with Docker and configured to use a locally mounted model folder. source
- Developer interfaces
- The project provides bindings for Python, Java, C#, and JavaScript, among other languages. source
- Export formats
- TXT,SRT source
- Integrations
- The integrations page lists Asterisk, FreeSWITCH, Jigasi, and UniMRCP integrations. source
- Intended users
- The FAQ describes Vosk as a practical, plug-and-play speech recognition library for applications, while suggesting Kaldi for research. source
- Languages
- The toolkit supports more than 20 languages and dialects. source
- License
- The Vosk API GitHub repository identifies its license as Apache-2.0. source
- Model sizes
- Small models are typically around 50 MB and require about 300 MB of runtime memory; large server models can require up to 16 GB. source
- Offline use
- Vosk works offline, including on lightweight devices such as Raspberry Pi, Android, and iOS. source
- Platform limits
- The installation page says ARMv6 and Windows ARM64 are not supported, and iOS builds are available on request. source
- Product
- Vosk is a speech recognition toolkit for offline speech recognition. source
- Server protocols
- Vosk Server has implementations for WebSocket, gRPC, MQTT, and WebRTC. source
- Speaker identification
- Vosk supports speaker identification alongside speech recognition. source
- Speaker identification
- Yes source
- Streaming
- Vosk provides a streaming API for speech recognition. source
- Support
- The project directs users to GitHub issues, email, Telegram, and an invitation-only WeChat group for questions. source
- Supported systems
- The installation page lists Linux x86_64, Linux arm64, Raspberry Pi 3/4, macOS x86 and M1, and Windows x86 and 64 for its Python package. source
- Timestamp support
- Yes source
- Vocabulary
- Vosk allows vocabulary reconfiguration, and its models page says most small models support dynamic vocabulary changes. source
Checked against the maker's pages on 11 Oct 2026.