Maurizio Copperi

CELP Coding for high-quality speech at 8 kbit/s

A new speech coding technique at low bit-rate is presented in this paper. The coder is based upon a novel speech production model, independently developed by the authors [1,2] and by Atal and Schroeder [3,4], called CELP (Codebook Excited Linear Prediction). Differences exist between the two approaches, both in the strategy chosen to construct codebooks, and in the method to generate the innovation sequence. In this scheme, we split the incoming speech signal into two frequency bands in order to gain the benefits of the piecewise LP (Linear Prediction) approximation. Then, each residual signal is coded in blocks of 5-ms duration through an adaptive vector quantizer incorporating a noise shaping filter. Our results show that good quality speech can be obtained at 8 kbit/s.

Vector quantization and perceptual criteria for low-rate coding of speech

Considerable effort has been and is currently being concentrated on improving the speech quality at low and very low bit rates. Recently new models of LPC excitation have been devised, which are able to yield good quality speech by exploiting our knowledge of the human speech production and perception processes. Unfortunately, these models generally require too much computational load to be easily implemented on currently available hardware. This paper describes an efficient speech coder, capable of providing acceptable quality speech, within the limitations of both low bit rate (approximately 2.4 kbit/s) and real-time implementation. The coder is based upon pattern classification and cluster analysis with perceptually-meaningful error minimization criteria. Our main objective is improving the excitation representation in a linear predictive coding scheme and, hence, the subjective quality of synthesized speech signals.

Design of a 4.8/9.6 kbps baseband LPC coder using split-band and vector quantization

This paper presents our recent work on the development of a 4.8/9.6 kbps baseband LPC coder, using vector quantization in both vocal tract modelling and residual representation. The vector quantizing approach has been proved highly effective for achieving good quality in voice communications applications at low and very-low data rates. Here we focus on the most important aspects in designing a baseband coder, with the goal of maximizing the subjective quality. Results of some preliminary experiments, performed on a FPS-120 array processor connected to a mini PDP11/60, are described. Taped demonstrations will be played at the Conference.

9.6 kbit/s Piecewise LPC residual excited coder using multiple-stage vector quantization

This paper describes a speech coder which exploits the combination of piecewise LPC analysis, full residual excitation and vector quantization in order to yield very good quality at 9.6 kbit/s. The piecewise approximation of the speech spectrum permits a higher spectral accuracy than standard LPC and reduces the computational load by about 40%. The full residual excitation overcomes the disadvantages of the base-band model, previously used at this bitrate. Finally, the vector quantization approach permits a dramatic bit saving over scalar quantization for LPC parameter compression, and provides a better distortion performance in the residual representation.

Efficient excitation modeling in a low bit-rate CELP coder

The author describes an efficient model of the excitation signal of a CELP (code excited linear prediction) coder, capable of reducing significantly the corresponding bit rate while maintaining a good subjective quality. This topic is of paramount importance since, typically, in CELP coders the greater part of the available bit rate is allocated to encode the excitation signal. The model has been improved by resorting to (1) a dual loop algorithm to compute pitch lags; (2) a codebook of excitation vectors designed by a covering method; and (3) a combined backward/forward adaptive estimator to track the excitation gain. Numerical results and listening tests show that this excitation model overcomes the typical deficiencies of CELP coders at low bit rates. In particular, the proposed approach makes it possible to reduce the excitation rate of a stochastic CELP coder by 20% while at the same time improving the speech quality in voiced speech.<<ETX>>

Rule-based speech analysis and application of CELP coding

An approach is presented for efficiently encoding speech signals at low bit rates, by exploiting a combination of various speech analysis and compression techniques cooperating via a rule-based reasoning system. A front-end analyzer compresses speech events at nonuniformly spaced time intervals, by resorting to dynamic and static variable-frame-rate methods relevant to perception models. Then a codebook-excited-linear-predictive (CELP) coder performs a perceptually meaningful identification and quantization of the excitation parameters, to provide an optimal rendition of the original signal. This coding scheme can reduce the transmission rate down to 2.4-2.8 kb/s, while retaining a very good quality. The main applications are in voice response systems and voice mail.<<ETX>>

16 kbit/s Split-band APC coder using vector quantization and dynamic bit allocation

This paper describes a speech coder which exploits the combination of subband splitting, adaptive predictive coding, vector quantization and dynamic bit allocation in order to yield very good quality at 16 kbit/s. The coder has been evaluated by means of objective measures and informal listening tests. Some experiments have also been conducted with input signals degraded by acoustic background noise. Results confirm that the proposed coder is a viable approach for such applications as end-to-end transmission circuits, digital mobile radio and voice mail systems.

Broadcasting-quality transmission of audio signals at 64 kbps

Recently, attention has been paid to the problem of transmitting audio signals whose bandwidth is wider than 3.4 kHz, tipycally 7 kHz. In this paper, three techniques are examined which allow the transmission rate of 64 kbit/s to be met. Comparisons are carried out in terms of both objective measurements and informal listening tests. Possible application fields are tele- and video-conferencing, commentary channels and broadcasting, and high quality speech and music in the future Integrated Service Digital Network.

An adaptive delta modulator with noise spectral shaping

Noise spectral shaping and performance analysis are considered in the context of Adaptive Delta Modulators (ADMs). The goal pursued is to reduce the in-band distortion by redistributing the noise power over all the frequency band available, in a manner depending upon the feedback network. Since ADMs generally work at a sampling frequency much higher than the Nyquist rate, one could obtain quality improvements by shifting a large amount of quantizing noise outside the base-band. Simulation results reveal that the benefit actually achievable is limited, essentially because the noise shifting causes an increase of slope overload distortion. A suitable ADM scheme with switched narrow-band Noise Shaping filter, permits to mitigate this adverse phenomenon and to achieve a subjective quality improvement.

Feature extraction and product codes in vector excited coders

A novel approach to code speech signals at low bit rates, while retaining acceptable subjective quality, is presented in this paper. The system falls into the class of vector excited coders, and exploits benefits stemming from the spectral masking effect of the quantizing noise. Since a main disadvantage of conventional vector excited coders is the huge amount of computation associated with the selection of the optimum excitation signal, our primary goal is the development of an efficient and computationalIy simple codebook search method. We present an original scheme based upon a 3-level hierarchical tree vector quantizer.