Pocketsphinx-based Linux Voice Dictation Service
Note:
You likely want Listener2
This old repository includes the original PocketSphinx based implementation, the new repository includes the modern Deep Speech based engine.
The basic idea is to see if CMU Pocketsphinx (a BSD-licensed continuous voice recognition system) can be used to create a voice dictation service reasonable enough to be used to drive a programmer's IDE/editor; not to completely eliminate the use of hands, but to augment them such that most identifiers and common code constructs can be dictated without touching the keyboard.
While editing/navigation/command and control will likely be looked at some time, they aren't a priority. The basic approach should work for any dictation task, it is just that the first task I want to work on is the one for which I need a solution.
Big Caveats:
This is an early stage project
- See The TODO List
- At the moment you can use command-line tools to create a dictation context from your code repositories, start a qt-based Listener instance and have that dictate into an application that listens on DBus. I've got an Eric IDE plugin that works as a client for that DBus service.
This project is currently English-only (and English-US keyboard only). It would be nice to support other languages, but it is not a priority for me
This project is Linux-only.I am targetting modern (K)ubuntu desktops.
The use of pocketsphinx is actually somewhat incidental. While we are using pocketsphinx right now, we should be able to switch out the engine for another at some point with little change to the GUI and services. The reason I'm using Pocketsphinx is that it comes nicely packaged under Ubuntu and provides pre-packaged language models for English-US dictation.
Setup
This is still very much a programmers/contributors only project. That said, these instructions should get you a working setup on a Kubuntu 14.04 machine.
Dependencies:
$ apt-get install gstreamer0.10-pocketsphinx build-essential \
libsphinxbase1 sphinxbase-utils sphinxtrain \
pocketsphinx-hmm-en-hub4wsj pocketsphinx-utils \
espeak alsa-utils python-gst0.10 python-gobject-2 \
python-sqlite build-essential
# for the Qt-based GUI (Note: pyside is LGPL)
$ apt-get install python-pyside.qtcore python-pyside.qtwebkit python-jinja2
# for the Desktop service (uinput), currently unimplemented
$ apt-get install python-dbus
Listener is a python library (using setuptools), use:
$ git clone https://github.com/mcfletch/listener.git $ cd listener $ python2.7 setup.py develop --user
to install.
Executables
listener-context-from-project --context=<name> /path/to/project
Uses word extraction to generate a language model (from Python files) based on a git checkout. Creates a new context <name>. With this done you should be able to dictate code as in the project using listener-qt (note: currently you will not see the actual code when dictating, you will see commands such as "cap" and "no-space".
listener-qt --context=<name>
Launches the (not-very-useful) Qt Listener GUI. You can dictate and see the results of each dictation. You can also click a button to review the raw audio captured for each utterance. Runs the code from listener-pipe in a background thread and uses Qt messages to communicate. Eventually the code from listener-pipe should be moved to a DBus service.
Note: this will download a large language model on first run. Currently that's done before the GUI starts, so the process will just seem to hang.
listener-context-delete --context=<name>
Delete a context and all associated data.
listener-code-to-words-py *.py
Attempts to do a code-to-words translation (for building a language model) Writes .py.dictation files next to the source files for manual review in order to improve the translations.
listener-missing-words *.py
Performs code-to-words translations and reports all words/tokens not in the default dictionary. This is mostly for use in improving the code to words code.
listener-pipe
Attempt to setup a gstreamer pipeline using a downloaded language model that matches the hub4wsj package. The pipeline will store the raw audio of utterances into ~/.config/listener/default/recordings and print out the partial and final results to the console.
Note: this will download a large language model on first run. Currently that's done before anything else, so expect a hang.
listener-rawplay <filename>
Plays a raw audio file as output by the listener-pipe into the recording directory (to allow the user to review the content before adding it to their training data-set)
listener-arpa-guess <words>
Prints out the best-guess ARPABet definition for the incoming words, these are the things you need to add to a '.dict' file for pocketsphinx, generated by extracting correspondences between espeak and the CMU dictionary project data-files.
listener-uinput-device
Test case that tries to do a uinput "send keys like" operation, operates at the Linux kernel uinput level, so should work with any environment (in theory it could even work on a console, though I have not tried that).
Internal Utilities
These just modify (json) structures that are part of the code-base that provide lookup tables used by the code.
listener-uinput-rebuild-mapping
Rebuilds the mapping from character to keystrokes. Currently this just reads a kernel header and applies some hand-coded keyboard mappings for a US-english keyboard. Eventually should use users local xkb mappings (including compose keys) to properly map characters.
listener-ipa-arpa-statmap
Re-extract IPA -> ARPABet statistical map, should the algorithm be improved
License
Listener's code is licensed under the BSD license (as is Pocketsphinx). You have accepted the licenses for the Ubuntu/Debian packages used by installing them above.
Note that when combined with other software Listener may fall under more restrictive licenses.
© 2014, Copyright by VRPlumber Consulting Inc. and the Contributors; All Rights Reserved.
Permission to use, copy, modify, and distribute this software and its documentation for any purpose and without fee or royalty is hereby granted, provided that the above copyright notice appear in all copies and that both the copyright notice and this permission notice appear in supporting documentation or portions thereof, including modifications, that you make.
THE AUTHOR VRPlumber Consulting Inc. and the Contributors DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS, IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE!
Listener includes copies of:
jQuery v2.1.1 | (c) 2005, 2014 jQuery Foundation, Inc. | http://jquery.org/license
Pure v0.5.0 Copyright 2014 Yahoo! Inc. All rights reserved. Licensed under the BSD License. https://github.com/yui/pure/blob/master/LICENSE.md
normalize.css v1.1.3 | MIT License | http://git.io/normalize Copyright (c) Nicolas Gallagher and Jonathan Neal
- pysideqtsingleapplication | BSD 2-Clause License http://stackoverflow.com/questions/12712360/qtsingleapplication-for-pyside-or-pyqt
Listener will download the following software/models when run:
- CMU HUB4 Language Model -- which provides a few extra files that are needed to update/modify the acoustic model over the files distributed in the Ubuntu repository
- CMU CLM TK -- which provides the command line tools required to update a language model for use with Sphinx
