I've collected some more performance data for Sphinx-4 In particular, I've been comparing our Sphinx-4 performance against Sphinx 3.3 (the fast, large vocabulary recongizer written in C). One thing that is great about working on speech recognition systems is that they are pretty easy to compare. I just have to compare the accuracy (expressed as a word error rate - WER) and the speed expressed in terms of real time (RT). For both of these metrics, lower is better. This table - Sphinx 3.3 vs. Sphinx 4 Performance shows how well Sphinx-4 performs compared to Sphinx 3.3 for a wide range of vocabularies (from very small with 11 words, to very large with 60,000 words).

This table shows that for all tests except for the very large vocabulary test (hub4) we are running faster and more accurate than Sphinx 3.3. We are still working on improving the large vocabulary performance .. we are almost there.


Post a Comment:
Comments are closed for this entry.

This blog copyright 2010 by plamere