In order to recognize speech, Speech Recognition Engines require
two types of files: the first, called an Acoustic Model, is
created by taking a very large number of transcribed speech
recordings (called a Speech Corpus) and 'compiling' them into
statistical representations of the sounds that make up each word. The
second is a Grammar or Language Model. A Grammar is a
relatively small file containing sets of predefined combinations of
words. A Language Model is a much larger file containing the
probabilities of certain sequences of words.
Most Acoustic Models used by 'Open Source' Speech Recognition engines are 'closed source'. They do not give you access to the speech audio (the 'source') used to create the Acoustic Model. If they do give you access, there are usually licensing restrictions on the distribution of the 'source' (i.e. you can only use it for personal or research purposes).
The reason for this is because there is no free Speech Corpus in a form that can readily be used, or that is large enough, to create good quality Acoustic Models for Speech Recognition Engines. Although there are a few instances of small FOSS speech corpora that could be used to create acoustic models, the vast majority of corpora (especially large corpora best suited to building good acoustic models) must be purchased under restrictive licenses.
As a result, Open Source projects that want to distribute their code freely must purchase restrictively licensed Speech Copora that limit distribution of the 'source' speech audio, but allow them to distribute any Acoustic Models they create.
VoxForge will address this problem by providing all Acoustic Models and their 'source' (i.e. transcribed speech audio) in GPL licensing format - which requires that the distribution of derivative works include access to the source used to create that work.
Every project that wants to build an acoustic model using a corpus with restrictive licensing must purchase their own copy. This is difficult for FOSS projects, which usually have no revenue. If a project does purchase such resources, the license restrictions will require them to keep the resources behind some kind of access barrier restricted to official project members. This takes away freedom and flexibility from end users and shrinks the pool of potential contributors to the project.
Most Open Source Speech Recognition Engines ("SRE"s) come with an Acoustic Model. However, these Acoustic Models are not interchangeable with other open source Speech Recognition engines. The way to address this problem is to provide the 'source code' for the Acoustic Models (i.e. the Speech Corpora used to create the Acoustic Models), and permit users to 'compile' it into Acoustic Models that can be used with the Open Source SRE of their choice.
VoxForge hopes to address this problem by creating a repository of
'source' speech audio and transcriptions, and by creating Acoustic
Models for each of the main Open Source Speech Recognition Engines
(such as Sphinx,
Julius, HTK
and ISIP)
.
Current Acoustic Models used by Open Source Speech Recognition Engines are not at the level of quality of Commercial Speech Recognition Engines.
VoxForge provides a central location that can collect GPL speech audio and transcriptions. As more speech audio data is collected, better Acoustic Models can be created, to the point that someday they will be comparable to Commercial Speech Recognition.
Most Open Source SREs are designed for command and control and IVR telephony type applications (e.g. Sphinx, HTK and ISIP). The Julius Speech Recognition Engine was designed for dictation applications, however the Julius distribution only includes Japanese Acoustic Models. But since it uses Acoustic Models trained using the HTK toolkit, it can also use Acoustic Models trained in other languages - like English. We just need hundreds of hours of transcribed speech audio to create English dictation Acoustic Models. This same audio data might also be used to permit the other open source Speech Recognition Engines to work in dictation applications.
Although the current focus of VoxForge is on Speech Recognition for IVR telephony applications or Command and Control applications on the desktop, when the amount of audio data collected reaches a certain threshold, this data can then be used in the creation of Acoustic Models for Open Source Dictation Applications.
Currently, you can easily create a single user Acoustic Model trained to recognize your own voice using open source speech recognition software - it just takes time and patience. VoxForge's main objective is to create multi-user Acoustic Models that can be used without training for:
To achieve this, VoxForge will serve as a repository for transcribed speech audio files that will be used to create continuously-improving Acoustic Models (as user contributions are merged into the VoxForge Multi-User Acoustic Model).
As
more and more
transcribed speech data is collected, the creation of single user
Acoustic Models will be made easier. This is because users will
be able to adapt
the VoxForge Multi-User Acoustic Model to recognize their voice, rather than to try to
create one from scratch. As even more
speech data is obtained, then
the VoxForge Multi-User Acoustic Model will be able to recognize speech without
needing to be adapted to a particular user's voice.
To
achieve this objective, we need your help. There are two ways to help:
- You create transcribed speech audio files using your own voice and submit them;
- We compile them into the VoxForge Speaker Independent Acoustic Model.
- You use our How-to's or Tutorials to learn how to create your own Acoustic Models and submit the speech audio and transcriptions to VoxForge;
- We compile them into the VoxForge Speaker Independent Acoustic Model.
We are currently focusing on collecting high quality audio data (48kHz/16-bit) and downsampling it for use with telephony IVR
(8kHz/16-bit) and desktop command and control applications (16kHz-48kHz/16-bit).
The advantage of collecting data with these higher frequency rates is that these audio files could also be used in the creation Acoustic Models for PC based dictation applications (note that robust statistical Language Models would also need to be created to work with these Acoustic Models).
More information on the creation of Acoustic Models.We believe that making Speech
Corpora available using an unrestrictive, BSD style license will not help the Open Source Community in this particular case. A BSD
style license permits users to distribute derivative works without
having to contribute the source of those modifications back to the
community. In our opinion, the Open Source Speech Recognition
community does not have the required threshold of users to create a
self-sustaining community using a BSD
style license. If there was a larger community, then there would
be a greater likelihood that a self-sustaining group would give back to
the community, even if not required to do so using a BSD style license.