Analysis-by-synthesis multimode harmonic speech coding at low bit rate.

摘要

To satisfy a growing demand for new speech communication technologies in the past decade, speech coders capable of delivering toll-quality speech at lower and lower rates have been developed at an astonishing pace. Many promising low-rate coding schemes have been proposed in recent years to bridge the gap between toll quality waveform coding at rates above 6.3 kb/s and moderate quality vocoding at rates around 2.4 kb/s. Specifically, at present, there is a surge of research and commercial interest to develop high quality speech coders operating at bit rates of 4 kb/s and below.; The main objective of this dissertation is to develop new techniques that allow a harmonic (sinusoidal) speech coder to overcome various limitations in order to achieve high quality at a rate of 4 kb/s or below.; In harmonic coders, speech is traditionally classified as voiced or unvoiced. Harmonic coding of speech uses the harmonic model for voiced speech and the noise model for stationary unvoiced speech. However, experimental evidence shows that the poor representation of the transitory speech segments results in a significant degradation in the reconstructed speech quality. In order to improve the speech model accuracy in non-stationary speech segments, this dissertation introduces a novel frequency domain speech model for transition coding. This model represents time-domain significant events (pulses) by using a generalized sinusoidal model. A closed-loop analysis-by-synthesis parameter estimation procedure was devised for the new transition speech model.; In order to improve the accuracy and robustness of parameter estimation in harmonic coders, we introduced a novel time domain analysis-by-synthesis parameter estimation method in the harmonic coding framework. In this method, we propose the use of a nonlinear time scale modification technique for overcoming the waveform matching obstacle in harmonic coders and thereby achieving time-domain closed-loop parameter estimation. The effectiveness of this method is demonstrated by a specific algorithm for pitch and class estimation.; In this research, we also conducted a comprehensive study of the quantization issue for the variable-dimension harmonic magnitude vector. We incorporated speech perceptual weighting with the non-square transform vector quantization (NSTVQ) for harmonic spectral magnitudes quantization. We showed that the weighted NSTVQ is a generalization of all existing linear transformbased variable-dimension vector quantization schemes. Within the framework provided by the weighted NSTVQ, two implementation schemes were proposed and compared with the respect to the computational complexity. We demonstrated that the weighted NSTVQ system has the ability to trade performance for complexity and memory storage by selecting different transforms and the length of fixed-dimension vectors.; In order to demonstrate the viability of the new techniques studied in this research, we designed a 4 kb/s analysis-by-synthesis multimode harmonic coder (AbS-MHC) employing the proposed frequency domain transition speech modeling technique, the proposed analysis-by-synthesis parameter estimation technique for pitch/class estimation, and the new variable-dimension vector quantization technique. The resulting AbS-MHC coder at a rate of 4 kb/s achieves a perceptual quality very similar to the G.723.1 coder at 6.3 kb/s as indicated by subjective test results.