Abstract:

Aiming for the problem that the harmonic model-based phase spectrum speech enhancement algorithm can only reconstruct the phase of voiced segment, which leads to speech distortion and auditory discontinuity, a new method to improve phase reconstruction by using signal-to-noise ratio (SNR) information and time-frequency features is proposed. First, the time-frequency characteristics related to phase distortion are introduced and the decision threshold is calculated. Then the phase deviation between noisy speech and clean speech is calculated by using the signal-to-noise ratio information. The two comparisons further estimate the phase of voiced and unvoiced speech, which can effectively improve the coherence of speech. Finally, the reconstructed phase is combined with the amplitude estimation of the improved binary hypothesis model and the speech enhancement is performed. Experiments on different speeches in different noise backgrounds show that phase deviation of the new algorithm is closer to the original signal. Compared with the comparison algorithm, the signal-to-noise ratio of the enhanced speech is increased by 2.39dB on average, and the perceptual evaluation of speech quality is increased by 0.12 on average, which effectively reduces the speech distortion and improves speech intelligibility.

Key words: phase reconstruction, SNR information, time-frequency characteristics, decision threshold, phase deviation