Construction: Adding sox noise reduction and possible speech optimizations when recording samples? At least trying to. Conversion to speex and back to wav with sox doing normalization has been added. No parameter to invoke this additional feature yet.

precise-listen may or may not be broken? In progress (driven by what we need for a customer).

Mycroft Precise

A lightweight, simple-to-use, RNN wake word listener.

Precise is a wake word listener. Like its name suggests, a wake word listener's job is to continually listen to sounds and speech around the device, and activate when the sounds or speech match a wake word. Unlike other machine learning hotword detection tools, Mycroft Precise is fully open source. Take a look at a comparison here.

Training Models

Communal models

Training takes lots of data. The Mycroft community is working together to jointly build datasets at https://home.mycroft.ai/#/precise. These datasets are used to build the models used by the Mark 1 and other mycroft-core based voice assistants. Please come and help make things better for everyone!

Train your own model

You can find info on training your own models here. It requires running through the Source Install instructions first.

Installation

If you just want to use Mycroft Precise for running models in your own application, you can use the binary install option. Note: This is only updated to the latest release, indicated by the latest commit on the master branch. If you want to train your own models or mess with the source code, you'll need to follow the Source Install instructions below.

Binary Install

First download precise-engine.tar.gz from the precise-data GitHub repo. This will get the latest stable version (the master branch). Note that this requires the models to be built the the same latest version in the master branch. Currently, we support both 64 bit desktops (x86_64) and the Raspberry Pi (armv7l).

Next, extract the tar to the folder of your choice. The following commands will work for the pi:

ARCH=armv7l
wget https://github.com/MycroftAI/precise-data/raw/dist/$ARCH/precise-engine.tar.gz
tar xvf precise-engine.tar.gz

Now, the Precise binary exists at precise-engine/precise-engine.

Next, install the Python wrapper with pip3 (or pip if you are on Python 2):

sudo pip3 install precise-runner

Finally, you can write your program, passing the location of the precise binary like shown:

#!/usr/bin/env python3

from precise_runner import PreciseEngine, PreciseRunner

engine = PreciseEngine('precise-engine/precise-engine', 'my_model_file.pb')
runner = PreciseRunner(engine, on_activation=lambda: print('hello'))

Source Install

Start out by cloning the repository:

git clone https://github.com/mycroftai/mycroft-precise
cd mycroft-precise

If you would like your models to run on an older version of precise, like the stable version the binary install uses, check out the master branch.

Next, install the necessary system dependencies. If you are on Ubuntu, this will be done automatically in the next step. Otherwise, feel free to submit a PR to support other operating systems. The dependencies are:

• python3-pip
• libopenblas-dev
• python3-scipy
• cython
• libhdf5-dev
• python3-h5py
• portaudio19-dev

After this, run the setup script:

./setup.sh

Finally, you can write your program and run it as follows:

source .venv/bin/activate  # Change the python environment to include precise library

Sample Python program:

#!/usr/bin/env python3

from precise_runner import PreciseEngine, PreciseRunner

engine = PreciseEngine('.venv/bin/precise-engine', 'my_model_file.pb')
runner = PreciseRunner(engine, on_activation=lambda: print('hello'))

In addition to the precise-engine executable, doing a Source Install gives you access to some other scripts. You can read more about them here. One of these executables, precise-listen, can be used to test a model using your microphone:

source .venv/bin/activate  # Gain access to precise-* executables
precise-listen my_model_file.pb

How it Works

At it's core, Precise uses just a single recurrent network, specifically a GRU. Everything else is just a matter of getting data into the right form.