STT Plugins

STT plugins are responsible for converting spoken audio into text


The base STT, this handles the audio in "batch mode" taking a complete audio file, and returning the complete transcription.

Each STT plugin class needs to define the execute() method taking two arguments:

  • audio (AudioData object) - the audio data to be transcribed.
  • lang (str) - optional - the BCP-47 language code

The bare minimum STT class will look something like

from ovos_plugin_manager.templates.stt import STT

class MySTT(STT):
    def execute(audio, language=None):
        # Handle audio data and return transcribed text
        return text


A more advanced STT class for streaming data to the STT. This will receive chunks of audio data as they become available and they are streamed to an STT engine.

The plugin author needs to implement the create_streaming_thread() method creating a thread for handling data sent through self.queue.

The thread this method creates should be based on the StreamThread class. handle_audio_data() method also needs to be implemented.

Entry point

To make the class detectable as an STT plugin, the package needs to provide an entry point under the mycroft.plugin.stt namespace.

      entry_points = {'mycroft.plugin.stt': 'example_stt = my_stt:mySTT'}

Where example_stt is is the STT module name for the plugin, my_stt is the Python module and mySTT is the class in the module to return.

List of STT plugins

Plugin Offline Streaming Type
ovos-stt-plugin-fasterwhisper ✔️ FOSS
ovos-stt-plugin-whispercpp ✔️ FOSS
ovos-stt-plugin-vosk ✔️ FOSS
ovos-stt-plugin-chromium API (free)
ovos-stt-plugin-http-server API (self hosted)
ovos-stt-plugin-pocketsphinx ✔️ FOSS
ovos-stt-azure-plugin API (key)
imagem neon-stt-plugin-google_cloud_streaming API (key)
imagem neon-stt-plugin-nemo ✔️ ✔️ FOSS
imagem neon-stt-plugin-nemo-remote ❌️ API (self hosted)

Standalone Usage

STT plugins can be used in your owm projects as follows

from speech_recognition import Recognizer, AudioFile

plug = STTPlug()

# verify lang is supported
lang = "en-us"
assert lang in plug.available_languages

# read file
with AudioFile("test.wav") as source:
    audio = Recognizer().record(source)

# transcribe AudioData object
transcript = plug.execute(audio, lang)

Plugin Template

from ovos_plugin_manager.templates.stt import STT

# base plugin class
class MySTTPlugin(STT):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        # read config settings for your plugin
        lm = self.config.get("language-model")
        hmm = self.config.get("acoustic-model")

    def execute(self, audio, language=None):
        # TODO - convert audio into text and return string
        transcript = "You said this"
        return transcript

    def available_languages(self):
        """Return languages supported by this STT implementation in this state
        This property should be overridden by the derived class to advertise
        what languages that engine supports.
            set: supported languages
        # TODO - what langs can this STT handle?
        return {"en-us", "es-es"}

# sample valid configurations per language
# "display_name" and "offline" provide metadata for UI
# "priority" is used to calculate position in selection dropdown 
#       0 - top, 100-bottom
# all other keys represent an example valid config for the plugin 
MySTTConfig = {
    lang: [{"lang": lang,
            "display_name": f"MySTT ({lang}",
            "priority": 70,
            "offline": True}]
    for lang in ["en-us", "es-es"]