Speech Intelligibility Enhancement Method

Speech intelligibility enhancement belongs to an emerging subfield of speech enhancement area. Comparing to other enhancement technology like background noise reduction and other interference reduction, the intelligibility enhancement focused more on the degraded speech signal itself. To the question of improving speech quality, the speech intelligibility enhancement provides a different perspective. It realizes the voice quality and robust through processing speech signal itself.The original intention of the paper is to solve the speech intelligibility problem occurring in a tandem coding system of 2.4kbit/s AMBE and 16kbit/s CVSD. The speech information meets loss when coding in a low bit rate. Especially for the transition part (also called non-stable part, which usually happens in the joint area of vowel and consonant or just consonant part of fast speaking).There are two explanations for the situation:One, the low bit coding can not capture the 'fast changing', the unstable properties of transition part. In the parameter estimation and modeling process, the information during the transition part is being lost or is distorted.During the voice production, the speaker grants the vowel part more energy than the transition part. Due to the listening masking effect, the transition part is missing in the energy of vowel. After coding, it appears deteriorate.So, the speech intelligibility enhancement of this paper is to provide a solution to the above two root causes. The solution includes three major parts: The transition part detection; the adaptive time modification of segments; the intensity scale. The research includes investigation on the available algorithm; integrating the algorithm into our system. During this period, the work will cover the evaluation of the algorithm and improvement before integration. For example, improving the segment detection algorithm based on the comb-shaping filter fundamental frequency estimation, using the harmonic match degree as the similarity function of segment detection, and the improved edition with energy-weighted; Separating the time modification technology of stable segment and unstable segment; the necessary of voice on time estimation; the evaluation and improvement of harmonic and none harmonic separation technology in HNM model.Last, after subject testing on the 5 CVC speech file and one normal speaking speech file. The integration is proved to be helpful for improving the speech intelligibility.