Vosk Speech Recognition Toolkit

Vosk is an offline, open-source speech recognition toolkit designed to scale from small devices (like Raspberry Pi or Android smartphones) to large clusters.

Key features include:

• Multi-language support: Supports 20+ languages and dialects (including English, German, French, Spanish, Chinese, Russian, and more).
• Small Model Footprint: Models are approximately 50 MB.
• Streaming API: Provides zero-latency response for continuous large vocabulary transcription.
• Advanced Features: Supports reconfigurable vocabulary and speaker identification.
• Use Cases: Ideal for chatbots, smart home appliances, virtual assistants, subtitle generation, and lecture/interview transcription.

Vosk is an offline, open-source speech recognition toolkit designed for high-performance transcription. It supports over 20 languages and dialects (including English, German, French, Spanish, Chinese, Russian, and more) and is optimized for a wide range of hardware, from small devices like Raspberry Pi and Android smartphones to large server clusters.

Key features include:

• Continuous large vocabulary transcription.
• Zero-latency response via a streaming API.
• Small model sizes (starting around 50 MB).
• Reconfigurable vocabulary and speaker identification.

Common use cases include chatbots, smart home appliances, virtual assistants, subtitle generation, and lecture/interview transcription.

The Vosk C# wrapper currently only supports the following architectures:

• win64 (Windows 64-bit)
• linux64 (Linux 64-bit)

Note that win32 is not supported due to library loading complexities. Support for OSX or Android is not currently implemented.

To build the project from source using Gradle, ensure you have the following prerequisites met:

1. Install libvosk: Either build from source or download the binaries.
2. Download a Vosk model: Obtain a model from the official Vosk models page.
3. Configure Environment Variables: Set the following variables so the build/tests can locate the necessary files:
   • VOSK_MODEL: The path to your downloaded Vosk model.
   • VOSK_PATH: The path to your libvosk installation.

Once configured, run the build command:

./gradlew build

To use the Vosk API in Java, you must use the provided Java bindings which rely on jnr-ffi.

1. Build the project: Use Gradle to build the demo project.
2. Configure the model: Download a Vosk speech model and unpack it into a folder named model within the root directory of the demo project.
3. Requirements: Ensure you are using a recent version of the JDK and Gradle.

# Example build command (assuming Gradle is installed)
./gradlew build

The Kotlin/Native target is disabled by default. To enable it for development, you must first ensure libvosk is installed in the proper UNIX directories as defined in src/nativeInterop/cinterop/libvosk.def.

To activate the native target, use one of the following methods:

• Set the environment variable NATIVE_EXPERIMENT=true.
• Modify build.gradle.kts by setting enableNative = true.

After syncing the project, run the following task to generate the Kotlin C bindings: cinteropLibvoskNative

# Method 1: Environment variable
export NATIVE_EXPERIMENT=true

# Method 2: Edit build.gradle.kts
# enableNative = true

To see a working example of the Vosk C# wrapper, navigate to the demo folder and execute the following command. Ensure you have already unpacked your Vosk model and the test audio file into the expected locations before running.

dotnet run

The Go batch processing example converts s16le converted audio files to text. It is designed to imitate the behavior of the Python test_gpu_batch.py example for batch processing.

Prerequisites:

• An audio file converted to s16le format.
• The libvosk.so library must be present in your system's library path (e.g., LD_LIBRARY_PATH).

Tested Environment:

• Ubuntu 24.04
• Go 1.18
• gcc-11
• NVIDIA driver 570.172.08

The vosk-js package is an FFI-NAPI wrapper for the Vosk library, allowing you to use Vosk's offline speech recognition capabilities in Node.js environments. The bindings largely follow the standard Vosk interface, though some methods may not be fully implemented in this wrapper.

For implementation details and specific usage patterns, refer to the demo folder in the repository.

For detailed installation instructions, code examples, and comprehensive documentation for all language bindings, visit the official Vosk website.

https://alphacephei.com/vosk