Sixth European Conference on Speech Communication and Technology
(EUROSPEECH'99)

Budapest, Hungary
September 5-9, 1999

Robust HMM to Variation of Noisy Environments Based on Variance Extension of Noise Models

Hiroshi Matsumoto, Hiroaki Ubukata

Faculty of Engineering, Shinshu University, Nagano-shi, Nagano, Japan

This paper addresses the problem of making the PMC-based HMM robust to variation of SNR. The method proposed consists of simply expanding the variace of cepstral coefficients for the noise model in the parallel model combination (PMC). The effect of this technique is examined through speaker independent isolated digit recognition tests using NOISEX-92 noise data. The results show that the variance expansion of several lower order cepstra extremely improves robustness to a wide range of SNR mismatch over the standard PMC. The appropriate expansion factor is such that the expanded variance of the zeroth cepstrum is around 20dB with respect to its geometric mean level.


Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Matsumoto, Hiroshi / Ubukata, Hiroaki (1999): "Robust HMM to variation of noisy environments based on variance extension of noise models", In EUROSPEECH'99, 2387-2390.