Spectral and Temporal Envelope Cues for Human and Automatic Speech Recognition in Noise.

Abstract

Acoustic features of speech include various spectral and temporal cues. It is known that temporal envelope plays a critical role for speech recognition by human listeners, while automated speech recognition (ASR) heavily relies on spectral analysis. This study compared sentence-recognition scores of humans and an ASR software, Dragon, when spectral and temporal-envelope cues were manipulated in background noise. Temporal fine structure of meaningful sentences was reduced by noise or tone vocoders. Three types of background noise were introduced: a white noise, a time-reversed multi-talker noise, and a fake-formant noise. Spectral information was manipulated by changing the number of frequency channels. With a 20-dB signal-to-noise ratio (SNR) and four vocoding channels, white noise had a stronger disruptive effect than the fake-formant noise. The same observation with 22 channels was made when SNR was lowered to 0 dB. In contrast, ASR was unable to function with four vocoding channels even with a 20-dB SNR. Its performance was least affected by white noise and most affected by the fake-formant noise. Increasing the number of channels, which improved the spectral resolution, generated non-monotonic behaviors for the ASR with white noise but not with colored noise. The ASR also showed highly improved performance with tone vocoders. It is possible that fake-formant noise affected the software's performance by disrupting spectral cues, whereas white noise affected performance by compromising speech segmentation. Overall, these results suggest that human listeners and ASR utilize different listening strategies in noise.