14th International Congress of Phonetic Sciences (ICPhS-14)
San Francisco, CA, USA
Line Spectrum Frequencies (LSFs) are parameters for the low bit-rate (LBR) speech coding. It is desirable to use LSFs for speech recognition without decoding the encoded speech into waveform. However, the performance of using LSFs in speech recognition degrades when the speech signal is contaminated by additive noise. This paper introduces a graphic representation of the effect of additive noise to LSFs. It shows that LSFs of noisy speech shift smoothly to LSFs of noise as SNR increases. Then an interpolation algorithm is developed to adapt the speech models in LSF domain. The concept of so-called LSF spectrum is applied to estimate the interpolation factors. Experiment has demonstrated the effectiveness of this proposed method.
Bibliographic reference. Yu, An-Tzyh / Wang, Hsiao-Chuan (1999): "Effect of noise on line spectrum frequency and a robust speech recognition method for the low bit-rate encoded speech", In ICPhS-14, 1697-1700.