Technology Overview
Mycroft is the name of a suite of software and hardware tools that use natural language processing and machine learning to provide an open source voice assistant.
Mycroft is modular. Some components can be easily 'swapped out' for others:
Wake Word detection
Speech to Text (STT)
Intent parser
A Wake Word is a phrase you use to tell Mycroft you're about to issue a command. By default, this is Hey Mycroft, but you can configure your own Wake Word in your Mycroft Home account.
There are two technologies that Mycroft.AI currently uses for Wake Word detection:
PocketSphinx: PocketSphinx is part of the broader CMUSphinx package, developed by Carnegie Mellon University. PocketSphinx is a lightweight speech recognition engine, specifically tuned for handheld and mobile devices.
Because PocketSphinx is trained on English speech, your Wake Word currently needs to be an English word, like Hello Mike, Hi there Mickey or Hey Mike. Wake Words in other languages, like Spanish, French or German, won't work as well.
Precise: Unlike PocketSphinx, which is based on Speech to Text technology, Precise is a neural network that is trained on audio data. It doesn't matter what words you want to use for your Wake Word. Instead, you train it on sounds. The downside is that Precise needs to be trained on your chosen Wake Word. Precise is the default Wake Word Listener for the "Hey Mycroft" wake word, PocketSphinx provides a fallback to this if Precise is unavailable.
Speech to Text (STT) software is used to take spoken words, and turn them into text phrases that can then be acted on.
We are working with Mozilla to build DeepSpeech. A fully open source STT engine, based on Baidu’s Deep Speech architecture and implemented with Google’s TensorFlow framework.
DeepSpeech is not yet ready for production use and Mycroft currently uses Google STT as the default STT engine.
Mycroft also supports other STT engines that can be configured using the Configuration Manager:
IBM Watson Speech to Text (IBM API key required)
wit.ai Speech to Text (wit.ai API key required)
An intent parser is software which identifies what the user's intent is based on their speech. An intent parser usually takes the output of a Speech to Text (STT) engine as an input.
For example, Julie Speaks the following to Mycroft: Hey Mycroft, tell me about the weather
Julie's intent is to find out about the weather (probably in her current location).
An intent parser can then match the intent with a suitable Skill to handle the intent.
Adapt intent parser: Adapt is the default intent parser for all Mycroft platforms. Adapt was developed by Mycroft and is available under an open source license.
Padatious: Padatious is a neural network based intent parser. Padatious is currently under active development by Mycroft and is available under an open source license. It is likely that some Mycroft platforms will switch to using Padatious in the future instead of Adapt.
Text to Speech (TTS) software takes written text, such as text files on a computer, and uses a voice to speak the text. Text to Speech can have different voices, depending on the TTS engine used.
Mimic: Mycroft's default local text to speech (TTS) engine, based on CMU's Flite (Festival Lite)
Mimic2: Mycroft's own cloud based text to speech (TTS) engine, based on Tacotron providing a much better voice quality.
In your home.mycroft.ai account, you can select voices from these as well as
Google TTS: you need to choose which voice to use
even more tts engines are available but require manual configuration.
The Mycroft middleware has two components:
Mycroft Core: this code, written in Python, is the core software that provides the 'glue' between other modules. Mycroft Core is available under an Apache 2.0 open source license.
Mycroft Home and Mycroft API: this is the platform where data on Users and Devices is held. This platform provides abstraction services, such as storing API keys that are used to access third-party services to provide Skill functionality. The code for this platform is available under an AGPL 3.0 open source license.
Mycroft Skills are like 'add-ons' or 'plugins' that provide additional functionality. Skills can be developed by Mycroft Developers, or by Community Developers, and vary in their functionality and maturity.
Mycroft Skills Kit (MSK) is a Python-based utility that has been created to make it easier for Skill Authors to create, test and submit Skills to the Skills Marketplace.
Mycroft Skills Manager (MSM) is a command line tool used to add, manage and remove Skills on any Mycroft installation.
Mycroft is designed to run on many different platforms. Each dedicated platform is called a device, these include:
Mark 1 - our first reference hardware device using a dedicated software image.
Mark 2 - our latest reference hardware device using a dedicated software image.
Picroft - any Raspberry Pi 3 or 4 that is running the Picroft software image.
The enclosure refers to the specific code that is required for that device. It might define unique functionality such as the eyes on the Mark 1, or a specific way of interacting with the hardware, such as controlling the volume levels at a hardware level via i2c.
