Introduction
EvalDictator team consists of many senior people from CMU, MERL,
NIH, Sun and ex-Dragon.
Our overall goal is to encourage a new generation of speech
recognition research and entrepreneurs by releasing state of the art open source
speech technology, and making massive amounts of speech data freely
available. Currently, speech recognition technology is only available
from a handful of very large companies. Licensing and customizing this
technology is very expensive and involves onerous and lengthy
negotiations. We believe that is a barrier to innovation in this
market.
In order to achieve these ends, we want to popularize speech
recognition technology by building open source applications. These tools will be
written in Java and will run on every major platform including
Windows, OSX and Linux. Our target is computer users who wish to enter text in
their native language, and prefer speech to the keyboard. These users
may be professionals who require hands free text entry. For example,
many Doctors prefer to enter reports via dictation. Another target is
users who find it difficult to type text in their native language. For
example, in many countries users must memorize many complicated key
combination to enter basic text. Large countries ( e.g. China,
Japan...) and countless smaller ethnic groups have this problem.
We also want to collect a huge corpus of speech and language data.
This data will be used for speech research, and to further improve the
EvalDictator technology. EvalDictator differs from standard Desktop
dictation products such as NaturallySpeaking and ViaVoice because
speaker adaptation and custom pronunciation dictionaries are performed
as an Internet service. The speech and language data used to build
custom models for individual users will be stored and used to improve
the general models for everyone. EvalDictator applications will
automatically update themselves as new models become available. The
result is that the accuracy of EvalDictator applications will
improve as more people use them.
We strongly believe that technology should be application driven. In
other words, we want to solve real problems using speech recognition
applications, and only extend the core technology as required by
those applications. The domain of speech recognition is far too big for
us to address all at once, so we want to focus on the tasks that will
make the technology popular and successful.
Finally, we are seeking support and funding. Collecting and storing
massive amounts of audio data will require a serious
infrastructure. We will need machines, storage, backup and a fat connection to the
Internet. We will also need people to maintain the system. Also, if we
afford to work on our software full time we can make much more rapid
progress. Therefore we are seeking customers who require a custom
solution that is not available from the existing vendors.
Therefore, we are currently building general tools, but looking for
vertical domains. We want to release focused applications that
solve a problem for some set of users. We hope to build our "Eval Empire" one
happy user group at a time.
EvalDictator source code is free and open source with an "Apache"
style license. The code may be used in proprietary products, even if the
products are not open source. The source code will be publicly
available on SourceForge.net. EvalDictator will contain no patented IP.
The acoustic and language models are also free, however, they need not
be built from publicly available speech data, or be built by publicly
available tools. In this way companies and researchers can donate
"eval" models to the project while keeping the data and tools proprietary.
Capabilities
Currently, there is a complete working prototype code named "HumbleBeginnings".
HumbleBeginnings is written in 100% Java, and consists of two parts: the Dictator and the Collector.
The Dictator is a desktop application that performs Dictation via automatic speech recognition. It is based on the Sphinx4 recognizer from Carnegie Mellon. Sphinx4 is a state-of-the-art large vocabulary recognizer. The accuracy is as good as anything, and the size and speed are sufficient for a modern PC. It currently supports English, Arabic and Mandarin using the Wall Street Journal and GALE models.
The Dictator UI is very simple. It is merely a text pane, where the user can insert text by speaking. The user simply selects the insert position with the mouse, clicks and holds while talking, and the text is inserted. The complete text is automatically placed on the clipboard. The user then goes to his/her favorite application (e.g. Word, Excel, OpenOffice), pastes the text and completes editing.
The Collector is a server based web application that creates custom dictionaries, language models and acoustic models. Users upload a collection of their own documents to the server. They are then asked to read a selection of their own text out load. The Collector then automatically creates a version of Dictator customized to their voice, language and vocabulary. Users may offer to donate this data to the project in order to improve future versions, otherwise it is deleted.
HumbleBeginnings is very early in its development. Dictator has very poor performance-- it is big, slow, inaccurate, and also has no command and control, no editing, no adaptation, and no vocabulary editor. The Collector is in development and not yet ready to process data. The team is discussing what techniques and features need to be implemented.
(NOTE: The links in this section point to local files created by
javadoc. If they are broken, please follow the instructions on
Creating Javadocs to create these links.)
Developed with IntelliJ