Threshold of degradation for frequency-distributed band-limited noise in continuous speech

Two series of perceptual experiments have been conducted to investigate the level of noise that acts as a threshold of degradation (TOD) when band-limited noise is added to speech such that the noise is either constrained to lie in one subband of the entire speech band or separately introduced into all subbands. In all cases the within-band signal-to-noise ratio is maintained invariant with time to model the encoding of subbands with the aid of adaptive quantizers. Below 1000 Hz, the TOD for 500 Hz wide bands is found to drop with frequency at roughly half the rate of the speech drop. Above 1000 Hz, the TOD for noise in individual subbands in only 10 dB and does not vary significantly with frequency. When noise is introduced simultaneously into all bands, the TOD is some 5 dB lower than when the noise is introduced into one band alone. Additional subdivision of the frequency bands for independent encoding, so as to increase the noise at TOD and thereby reduce the required SNR, is particularly effective at the lower speech frequencies, because it allows one to take advantage of both the rapidly dropping speech spectrum as well as the time incoherence of the speech energy variations within the various bands. The perceptual tolerance appears guided by the extent to which the noise can be masked by the speech energy present. Encoding techniques that take better advantage of this factor are likely to lead to transmission efficiencies while maintaining excellent speech quality.