SUMMARYThis paper discusses a noise suppressor for use in the standard AMR (adaptive multi-rate) voice codec in 3GPP (3rd Generation Partnership Project). The performance standards (TS 26.077) for 3GPP noise suppressors require both high quality and high noise suppression performance with limited computational complexity. This paper is based on the spectral subtraction method, and discusses the configuration in which spectral subtraction and amplitude suppression are mutually controlled according to the state of the input signal. The improved estimation of the subband S/N (signal-to-noise ratio) for use in the control of spectral subtraction and amplitude suppression is also discussed. The proposed method is evaluated through simulation experiments, and its effectiveness is verified. Subjective evaluation tests (evaluation by third parties) and objective evaluation tests were performed in accordance with 3GPP TS 26.077, and all 3GPP performance requirement standards were met. The system obtained the first performance endorsement from 3GPP in March 2002.

SUMMARYIn this paper, the authors propose three new excitation codebook search methods for suppressing the selection of an excitation code vector that has a high likelihood of producing pulse-like degraded sound in a low-bit-rate CELP coder that uses a pulse-like excitation code vector. The proposed methods are (1) adaptive weighting for coding distortion based on the state of the input speech, (2) weighting for coding distortion due to amplitude distribution distortion in the time direction, and (3) setting an upper limit for the coding distortion rate for pulse-like excitation codebook searches. Although an excitation code vector with large coding distortion is selected if these proposed methods are set from the coding distortion minimization standard that is used in a conventional CELP excitation code vector search, there is an improvement in the quality that is perceived by the sense of hearing. The authors applied the proposed methods to a 4-kbps CELP coder to verify their effectiveness according to subjective quality evaluation experiments. The results verified the effectiveness of the proposed methods by showing that the introduction of the proposed methods improved the MOS value by approximately 0.3 for clean speech and by approximately 0.4 for noise-added speech.