10.5.1 Environmental Variation and Noise

Environmental variation has received the most attention from the speech literature, and a number of techniques have been suggested for dealing with environmental noise. Spectral subtraction, for example, is used to combat additive noise. Additive noise is noise from external sound sources like engines or wind or fridges that is relatively constant and can be modeled as a noise signal that is just added in the time domain to the speech waveform to produce the observed signal. In spectral subtraction, we estimate the average noise during non-speech regions and then subtract this average value from the speech signal. Interestingly, speakers often compensate for high background noise levels by increasing their amplitude, F0, and formant frequencies. This change in speech production due to noise is called the Lombard effect, named for Etienne Lombard who first described it in 1911 (Junqua, 1993).

Other noise robustness techniques like cepstral mean normalization are used to deal with convolutional noise, noise introduced by channel characteristics like different microphones. Here we compute the average of the cepstrum over time and subtract it from each frame; the average cepstrum models the fixed spectral characteristics of the microphone and the room acoustics (Atal, 1974).

Finally, some kinds of short non-verbal sounds like coughs, loud breathing, and throat clearing, or environmental sounds like beeps, telephone rings, and door slams, can be modeled explicitly. For each of these non-verbal sounds, we create a special phone and add to the lexicon a word consisting only of that phone. We can then use normal Baum-Welch training to train these phones just by modifying the training data transcripts to include labels for these new non-verbal 'words' (Ward, 1989). These words also need to be added to the language model; often by just allowing them to appear in between any word.