Name
Abstract | ||
|---|---|---|
This paper focuses on the estimation of short-term linear predictive parameters from noisy speech and their subsequent use in wave- form enhancement schemes. We use a-priori information in the form of trained codebooks of speech and noise linear predictive coefficients. The excitation variances of speech and noise are de- termined through the optimization of a criterion that finds the best fit between the noisy observation and the model represented by the two codebooks. Improved estimation accuracy and reduced com- putational complexity result from classifying the noise and using small noise codebooks, one for each noise class. For each segment of noisy speech, the classification scheme selects a particular noise codebook. Experimental results show good performance, especially under non-stationary noise conditions. Listening tests confirm that the new method outperforms conventional speech enhancement sys- tems. the noise obtained from multiple frames. We propose a maximum- likelihood classification scheme to select a single codebook. The system is easily extendable to different noise types with the addition of the appropriate noise codebook. Since the individual noise code- books are typically smaller than a single noise codebook trained for all noise types, computational complexity is reduced. Smaller codebooks also mean that the advantage due to a-priori information is retained. Large codebooks, trained on different noise types lose this advantage to some extent, since with increasing size they pro- vide a less restrictive representation of the noise parameter space. In this case, the speech and noise codebook entries that maximize the likelihood score in the joint codebook search may no longer be the speech and noise codebook entries that represent the underlying speech and noise data. A similar classified scheme is used in (9) in the context of hidden Markov model (HMM) based enhancement using multiple noise HMMs, where a single noise HMM is selected during peri- ods of non-speech activity. The selected noise HMM is used until the next occurrence of non-speech activity when a new selection is made. In the classified scheme proposed in this paper, we per- form a classification for each frame of noisy speech using an aver- age estimate of the noise obtained from the observation. A more important difference is that in the method proposed here, the ex- citation variances are computed for each frame to better deal with non-stationary noise. |
Year | Venue | Field |
|---|---|---|
2004 | European Signal Processing Conference | Speech enhancement,Signal processing,Value noise,Pattern recognition,Noise measurement,Computer science,A priori and a posteriori,Speech recognition,Artificial intelligence,Linear predictive coding,Codebook,Computational complexity theory |
DocType | ISBN | Citations |
Conference | 978-320-0001-65-7 | 0 |
PageRank | References | Authors |
0.34 | 7 | 3 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Sriram Srinivasan | 1 | 379 | 27.92 |
| Jonas Samuelsson | 2 | 165 | 11.19 |
| W. Bastiaan Kleijn | 3 | 1110 | 106.92 |