speech_to_text_linux

A Linux implementation of the speech_to_text plugin.

Linux does not ship a system speech recognition engine, so this package performs recognition offline using the open-source Vosk toolkit. Microphone audio is captured through PulseAudio as 16 kHz mono PCM and streamed to Vosk, which produces partial and final results in real time. It means we do not need to use any cloud provider, or expensive AI-based compute.

Vosk is licensed by Apache 2.0.

Usage

This package is a platform implementation of speech_to_text:

dependencies:
  speech_to_text: ^7.4.0
  speech_to_text_linux: ^0.0.1

No other code changes are needed, once the dependency is present the plugin registers itself as the Linux implementation and the normal speech_to_text API works :)

See speech_to_text.

Build dependencies

Install the development packages required to build the native plugin:

# Debian / Ubuntu
sudo apt-get install libgtk-3-dev libpulse-dev unzip

# Fedora
sudo dnf install gtk3-devel pulseaudio-libs-devel unzip

# Arch Linux (based)
sudo pacman -S gtk3 libpulse base-devel unzip

Vosk library and model

No manual Vosk setup is required. At build time the plugin looks for a local libvosk.so / vosk_api.h and, if none is found, automatically downloads the prebuilt Vosk SDK (x86_64, aarch64, armv7l or riscv64) from the official releases and caches it under ~/.cache/speech_to_text_linux/ (unzip must be installed for the extraction). The resolved libvosk.so is bundled into the application automatically. At runtime the plugin downloads and caches the small en-US model on first initialize.

Bundling your own Vosk library

You only need to provide Vosk yourself if you want to override the defaults, for example:

  • Bundle or pin a specific libvosk.so (a custom build, or a version other than the auto-downloaded one) - install it to a standard prefix (/usr/local/lib,/usr/local/include) or set VOSK_DIR to the folder containing libvosk.so and vosk_api.h before building. A locally found library always wins over the auto-download, and it is the copy that gets bundled into your app:

    export VOSK_DIR=/opt/vosk

    Alternatively pass -DVOSK_VERSION=<version> to fetch a different prebuilt release, or -DVOSK_AUTO_DOWNLOAD=OFF to forbid downloading entirely (the build the fails unless a local copy is found).

  • Unsupported architecture - the auto-download only covers the architectures listed above; on anything else the build stops and asks for VOSK_DIR.

  • Offline / hermetic CI builds - with no network at build time, pre-provision the SDK and point VOSK_DIR at it (or pre-populate the cache directory).

Bundling your own model

By default the small en-US model is fetched and cached at runtime. To ship or select a different model, pass platform options to initialize:

await speech.initialize(
  options: [
    // Use a model you bundle/install yourself; disables the runtime download. you can link this to application storage & such.
    SpeechConfigOption('linux', 'modelPath', '/opt/vosk/vosk-model-small-en-us-0.15'),

    // ...or keep auto-download but fetch a different model:
    SpeechConfigOption('linux', 'modelName', 'vosk-model-small-de-0.15'),
    SpeechConfigOption('linux', 'modelUrl',
         'https://alphacephei.com/vosk/models/vosk-model-small-de-0.15.zip'),

    // ...or disable downloading entirely (initialize fails without a modelPath):
    SpeechConfigOption('linux', 'autoDownloadModel', false),
  ],
);

Notes / Odd Behaviour

  • Microphone access is unrestricted under a normal desktop session, so hasPermission returns true. Sandboxed packaging (Flatpak/Snap) may still require granting microphone access. Which may need to be done via another API? Will need to have a look.
  • Vosk does not expose a locale enumeration API; locales() reports the locale derived from the configured model (defaulting to en_US).

TODO: make it so it gets locale from inputted URL/model, since a different locale can be used per modal? maybe make list, not sure..

Libraries

speech_to_text_linux