The CMU Sphinx Group Open Source Speech Recognition Engines

The information presented here provides information about current uses of the severla versions of Sphinx. These may be used as guidelines, but keep in mind that there is no definite answer.

PocketSphinx is CMU's fastest speech recognition system. It uses Hidden Markov Models (HMM) with semi-continuous output probability density functions (PDF). Even though it is not as accurate as Sphinx-3 or Sphinx-4, it runs at real time, and therefore it is a good choice for live applications. You can find further documentation about PocketSphinx in the release documentation, or at the online documentation.

Sphinx-3 is CMU's state-of-the-art large vocabulary speech recognition system. It uses Hidden Markov Models (HMM) with continuous output probability density functions (PDF). It supports several modes of operation. The more accurate mode, known as the "flat decoder", is descended from the original Sphinx-3 release (still available for reference purposes at https://cmusphinx.svn.sourceforge.net/svnroot/cmusphinx/trunk/archive_s3/s3). The faster mode, known as the "tree decoder", was developed separately. The two decoders were merged in Sphinx 3.5, though the flat decoder was not fully functional until Sphinx 3.7. Further documentation can be found in the release documentation, or at the online documentation.

For further detail, please check the Sphinx-4 page.

Sphinx-2 is a fast speech recognition system, the predecessor of PocketSphinx. It is not being actively developed at this time, but is still widely used in interactive applications. It uses Hidden Markov Models (HMM) with semi-continuous output probability density functions (PDF). Even though it is not as accurate as Sphinx-3 or Sphinx-4, it runs at real time, and therefore it is a good choice for live applications. You can find further documentation about Sphinx-2 in the release documentation, or at the online documentation.

We have some regression tests comparing sphinx4 to s3 (flat decoder) and to s3.3 (fast decoder) in several different tasks, ranging from digits only to medium-large vocab. s3 (flat decoder) is often the most accurate, but Sphinx4 is faster and more accurate than sphinx3.3 in some of these tests.

If you're familiar with ARPA evaluations (using databases available via the Linguistic Data Consortium (LDC) at U.Penn), you can find WSJ 5k and RM1 results in the "Medium vocab" table, and WSJ 20k in the "Large vocab" table.

The decision about which version to use depends on how familiar you are with C (sphinx3) or Java (sphinx4), and how easy it is to integrate these into your system.