Name
Abstract | ||
|---|---|---|
Gaussian Mixture Model (GMM) computation is known to be one of the most computation-intensive components of speech recognition. In our previous work, context-independent model based GMM selection (CIGMMS) was found to be an effective way to reduce the cost of GMM computation without significant loss in recognition accuracy. In this work, we propose three methods to further improve the performance of CIGMMS. Each method brings an additional 5-10% relative speed improvement, with a cumulative improvement up to 37% on some tasks. Detailed analysis and experimental results on three corpora are presented. Most modern large vocabulary continuous speech recognition (LVCSR) systems rely on continuous density hidden Markov models (HMMs) for acoustic modeling. They consist of thousands of HMM states, each state modeled by a separate Gaussian mixture model (GMM), consisting of tens of multi- dimensional Gaussian densities. The function of the GMM computation component in an LVCSR system is to provide HMM state scores for the search routine. Given the large number of GMMs in the acoustic model, this computation is expensive unless done intelligently. In the past, several attempts have been made to speed up GMM computation (3- 6). The key to most of these speed-up techniques is being able to intelligently ignore some parts of the computation without significant loss of accuracy. As no single technique is sufficient to provide enough speed gain, it becomes necessary to apply several techniques simultaneously. A major concern in applying these methods in a practical system is that the individual techniques are not usually orthogonal to each other; gains from each do not necessarily accumulate. Researchers can have a hard time combining them together effectively. |
Year | Venue | Keywords |
|---|---|---|
2005 | INTERSPEECH | gaussian mixture model,speech recognition,cumulant |
Field | DocType | Citations |
Pattern recognition,Computer science,Speech recognition,Artificial intelligence,Mixture model,Computation | Conference | 3 |
PageRank | References | Authors |
1.10 | 6 | 3 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Arthur Chan | 1 | 239 | 15.28 |
| Mosur Ravishankar | 2 | 250 | 19.46 |
| Alex Rudnicky | 3 | 1726 | 202.34 |