Digital Signal Processing Reference
In-Depth Information
Figure 9-6.
Optimized insertion penalty as a function of prosodic stream weight
in white
noise condition.
errors (ms) were measured by comparing the detected boundary locations in
noise-added utterances with that in clean utterances. The mean digit boundary
detection error rate was reduced by 23.2% for 10dB SNR utterances and 52.2%
for 5dB SNR utterances using the
SP-HMM-DV.
These results indicate the ef-
fectiveness of prosodic information in digit boundary detection.
5.
CONCLUSIONS
This paper has proposed an extraction method using the Hough transform
and a new speech recognition method using syllable HMMs utilizing both seg-
mental and prosodic information. Both methods were confirmed to be robust in
various noise conditions. The prosodic information is effective in digit boundary
detection and consequently improves connected digit recognition performance
under noise. Future works include combination of our method with model adap-
tation or feature normalization techniques for noise effects and evaluation using
more general recognition tasks.
