A new technique for the analysis of speech, the perceptual linear predictive (PLP) technique, is presented and examined. This technique uses three concepts from the psychophysics of hearing to derive an estimate of the auditory spectrum: (1) the critical-band spectral resolution, (2) the equal-loudness curve, and (3) the intensity-loudness power law. The auditory spectrum is then approximated by an autoregressive all-pole model. A 5th-order all-pole model is effective in suppressing speaker-dependent details of the auditory spectrum. In comparison with conventional linear predictive (LP) analysis, PLP analysis is more consistent with human hearing. The effective second formant F2' and the 3.5-Bark spectral-peak integration theories of vowel perception are well accounted for. PLP analysis is computationally efficient and yields a low-dimensional representation of speech. These properties are found to be useful in speaker-independent automatic-speech recognition.

Perceptual linear predictive (PLP) analysis of speech

A summary of many of the new techniques developed in the last two decades for spectrum analysis of discrete time series is presented in this tutorial. An examination of the underlying time series model assumed by each technique serves as the common basis for understanding the differences among the various spectrum analysis approaches. Techniques discussed include the classical periodogram, classical Blackman-Tukey, autoregressive (maximum entropy), moving average, autotegressive-moving average, maximum likelihood, Prony, and Pisarenko methods. A summary table in the text provides a concise overview for all methods, including key references and appropriate equations for computation of each spectral estimate.

Spectrum analysis&#8212;A modern perspective

Quantization, the process of approximating continuous-amplitude signals by digital (discrete-amplitude) signals, is an important aspect of data compression or coding, the field concerned with the reduction of the number of bits necessary to transmit or store analog data, subject to a distortion or fidelity criterion. The independent quantization of each signal value or parameter is termed scalar quantization, while the joint quantization of a block of parameters is termed block or vector quantization. This tutorial review presents the basic concepts employed in vector quantization and gives a realistic assessment of its benefits and costs when compared to scalar quantization. Vector quantization is presented as a process of redundancy removal that makes effective use of four interrelated properties of vector parameters: linear dependency (correlation), nonlinear dependency, shape of the probability density function (pdf), and vector dimensionality itself. In contrast, scalar quantization can utilize effectively only linear dependency and pdf shape. The basic concepts are illustrated by means of simple examples and the theoretical limits of vector quantizer performance are reviewed, based on results from rate-distortion theory. Practical issues relating to quantizer design, implementation, and performance in actual applications are explored. While many of the methods presented are quite general and can be used for the coding of arbitrary signals, this paper focuses primarily on the coding of speech signals and parameters.

Vector quantization in speech coding

With the continuing growth of modern communications technology, demand for image transmission and storage is increasing rapidly. Advances in computer technology for mass storage and digital processing have paved the way for implementing advanced data compression techniques to improve the efficiency of transmission and storage of images. In this paper a large variety of algorithms for image data compression are considered. Starting with simple techniques of sampling and pulse code modulation (PCM), state of the art algorithms for two-dimensional data transmission are reviewed. Topics covered include differential PCM (DPCM) and predictive coding, transform coding, hybrid coding, interframe coding, adaptive techniques, and applications. Effects of channel errors and other miscellaneous related topics are also considered. While most of the examples and image models have been specialized for visual images, the techniques discussed here could be easily adapted more generally for multidimensional data compression. Our emphasis here is on fundamentals of the various techniques. A comprehensive bibliography with comments is included for a reader interested in further details of the theoretical and experimental results discussed here.

Image data compression: A review

A tutorial on signal processing in state-of-the-art speech recognition systems is presented, reviewing those techniques most commonly used. The four basic operations of signal modeling, i.e. spectral shaping, spectral analysis, parametric transformation, and statistical modeling, are discussed. Three important trends that have developed in the last five years in speech recognition are examined. First, heterogeneous parameter sets that mix absolute spectral information with dynamic, or time-derivative, spectral information, have become common. Second, similarity transform techniques, often used to normalize and decorrelate parameters in some computationally inexpensive way, have become popular. Third, the signal parameter estimation problem has merged with the speech recognition process so that more sophisticated statistical models of the signal's spectrum can be estimated in a closed-loop manner. The signal processing components of these algorithms are reviewed. >

/pdf/signal-modeling-techniques-in-speech-recognition-2s2amcdjv6.pdf

Signal modeling techniques in speech recognition

Several alternate sets of parameters that represent the linear predictor are investigated as transmission parameters for linear predictive speech compression systems Although each of these sets provides equivalent information about the linear predictor, their properties under quantization are different The results of a comparative study of the various parameter sets are reported Specifically it is concluded that the reflection coefficients are the best set for use as transmission parameters A more detailed investigation of the reflection coefficients is then carried out using a spectral sensitivity measure A method of optimally quantizing the reflection coefficients is derived using a minimax spectral error criterion and the sensitivity analysis The method consists of transforming the reflection coefficients to the so-called log area ratios and linearly quantizing them A qualitative study on changes in pole locations due to quantization serves to corroborate the use of this optimal quantization An optimal bit allocation strategy for the transmission parameters is also presented The use of another spectral sensitivity measure renders logarithms of the ratios of normalized errors associated with linear predictors of successive orders as the optimal quantization parameters Informal listening tests indicate that the use of log area ratios for quantization leads to better synthesis than the use of log error ratios

Quantization properties of transmission parameters in linear predictive systems

In any vocoder, the process of parameter quantization is crucial for reducing the transmission rate while maintaining the quality of the synthesized speech. We have observed that, for linear predictive vocoders, the short‐time spectral dynamic range of the speech signal is the single most important factor that affects the quantization properties of transmission parameters, irrespective of the particular choice of parameters. We describe two preprocessing methods by which the spectral dynamic range is reduced, thereby improving quantization properties. In the first method, adaptive optimal (linear predictive) preemphasis is applied either directly to the speech signal or more easily to the autocorrelation coefficients through convolution. Experiments indicate that an appropriate set of preemphasis filters can be pre‐selected; for any speech segment being analyzed, the filter from this set that is closest to the optimal choice can be used for preemphasis. The advantages of such a scheme as well as the effect of first‐ and second‐order preemphasis on quantization and speech quality are discussed. The second method of preprocessing (the SIGMA method) involves multiplying the impulse response of the inverse linear prediction filter by an exponential window. Some preliminary results obtained using this method are reported.

Adaptive Preprocessing for Linear Predictive Speech Compression Systems

: This report describes work in developing a linear predictive speech compression system that transmits high quality speech at low bit rates. The authors have developed several methods for reducing the redundancy in the speech signal without sacrificing speech quality. Included among these methods are preemphasis of the incoming speech signal, adaptive optimal selection of predictor order, optimal selection and quantization of transmission parameters, variable frame rate transmission, optimal encoding, and improved synthesis methodology. When all of these were incorporated a floating point simulation of a pitch-excited linear predictive vocoder, synthesized speech with high quality at average transmission rates as 1500 bps was obtained.

Natural Communication with Computers. Volume 2. Speech Compression Research at BBN

We consider pitch‐excited linear predictive vocoders, and present approaches to achieve minimal transmission rates while maintaining good‐quality speech. Our contributions deal with quantization and encoding of predictor parameters before transmission, and their interpolation before synthesis. For purposes of quantization and transmission we have found the so‐called log area ratios to be optimal in that they preserve filter stability under quantization, and they possess uniform spectral sensitivity characteristics (Makhoul and Viswanathan, BBN Report No. 2800). The quantized parameters are encoded using (i) a variable‐rate scheme that transmits parameters only when speech characteristics have sufficiently changed, and (ii) a variable‐wordlength scheme (Huffman code) that uses parameter statistics to code parameters with variable numbers of bits. At the receiver, the synthesizer parameters are reset at the corresponding times at which they were extracted during the analysis. In addition, between every two ...

Towards a minimally redundant linear predictive vocoder

: This report describes our work in the past three years on data compression and quality evaluation of digital speech We developed and implemented linear predictive coding (LPC) techniques with the overall objective of digitally transmitting high quality speech at the lowest possible average data rates over packet-switched communication media Major techniques reported include: covariance lattice method of linear prediction analysis, adaptive lattice methods, linear predictive spectral warping, improved quantization of LPC parameters, variable frame rate transmission of LPC parameters based on a functional perceptual model of speech, and a mixed-source model for LPC synthesizer to produce more natural-sounding speech Also, we developed a reliable method for measuring subjective speech quality This method was employed to formally demonstrate the quality improvements provided by our speech analysis/synthesis techniques as well as for studying speech quality as a function of LPC parameters As subjective procedures are generally expensive and time-consuming, we developed and tested several objective procedures for speech quality evaluation The results from these objective procedures were found to be highly correlated to the corresponding subjective quality judgments Another highlight of our work is the development of a speech processing computer facility with the ultimate goal of transmitting narrowband speech in real time over the ARPA Network

R. Viswanathan

Papers

Quantization properties of transmission parameters in linear predictive systems

Adaptive Preprocessing for Linear Predictive Speech Compression Systems

Natural Communication with Computers. Volume 2. Speech Compression Research at BBN

Towards a minimally redundant linear predictive vocoder

Speech compression and evaluation