Disclaimer : Information provided here may not represent the standpoints of Carnigie Mellon University or CMU Sphinx Group.
Most of our effort in CMU was put into improving core speech recognition technology, we also realized that in terms of software engineering, we have a lot of room for improvement. This is why we have this page. I guess you can say it is my list of imagination because most of them are pretty hard to achieve in general.
If you are interested in any of the following projects, please send mail to me at archan at cs dot cmu dot edu
Usage of Sphinx 2, Sphinx 3, Sphinx 3.4 and Sphinx 4 has very steep learning curve. Major reason is that there is large amount of command line arguments (>50) in each of these recognizers. We are calling for programmers to help us to write a graphical user interface, which aid the user to decide 1, which parameter(s) they should modify, 2, how they should modify them (e.g. The range of values of a parameter.)
The target language is not yet decided, portable language for the task. Currently, perl/tk, C++ with wx and Java are three candidates we are considering because they are all reasonably portable across different platforms.
HMM-based speech recognition has come to a stage where file format matters. Common softwares such as HTK, Sphinx and ISIP Speech recognizer all assumes slightly different formats. In this project, we wish to build a converter such that models built by one trainer can be freely used by other decoders. We also hope to
New at 20050503! Someone has already taken this part and we will work in the next few months, we will make a formal annoucement later.
The target language is C or perl. We hope that both interfaces exist at the same time.
David Huggins of Cepstral Inc. has built a perl module to wrap Sphinx II. We need similar interface for Sphinx III and Sphinx IV.
The target language would be in perl.
Sphinx was mainly developed by researchers and developers. It was always a single thread operation mode. However, in practice, there are many situations where Sphinx is needed to be used in a thread, we want to evaluate whether sphinx can do it. Also, if it need to be MT-safe, how much effort we need to put in it.
New at 20050115! Someone has already taken this part and we will work in the next few months, we will make a formal annoucement later.
Sphinx development started 16-17 years ago where people is not aware of issue in computer internationalization. We hope that Sphinx can eventually be used to be input with unicode-text and output with unicode-text.
The use of open source speech recognizer was hindered by the fact that is no good opens source speech browser. Speechworks (now part of Scansofts) and CMU have developed OpenVXML and OpenSALT as a resource browser. We hope to make an effort for building a more generic interface for dealing with multiple standards in the field.
Sphinx is portable in various platforms, however, it is not well tested in platforms such as BeOS, ALSA and OS2. We hope that we can make Sphinx to be useful in those platform as well.
New at 20051203! David Huggins-Daines has created a version of Sphinx 2 which works at a Sharp Zaurus. See here
Imagine a speech recognizer which is ubiquitous and can follow you to everywhere of the world. Many commercial speech recognizer can already do it. Sphinx is possibly a very good candidate to do the same. This project will require certain level of understanding of speech recognizer and programming using integers.
New at 20050503! Currently, sphinx 3 is configured using simple command-line and configuration becomes pretty difficult in general. Though the code is developed such that conceptually it has a hierachial, it lacks of similar interface right now. XML seems to be a very good candidate for the task.
This is half-joking :-) Someone suggested to me that Sphinx may need a song. Well, that is actually not a very bad idea. I don't know music at all. So, just tell me what you think is good. :-)