LPC Coefficients Research Articles

An integrated voice analyzer for evaluating pathological voice

We have developed an integrated voice analyzer (IVA) using DSP (TMS32010) to achieve high-speed acoustic analyses of a sustained vowel. A personal computer connected to the IVA is employed as the man-machine interface as well as to display and save various analysis results. The IVA has many functions, such as detecting vocal noise by an adaptive comb filter, measuring period and amplitude perturbation quotients, calculating DFT spectra and LPC coefficients, etc. The results are stored into a disk file to do statistical analysis or to perform further analyses. This system provides an interactive acoustic analysis method for quantitative evaluations of the pathological voice arising in various voice research and clinical areas. We describe an application of the IVA for the clinical screening system for pathological voices, where hoarseness grade (HG) of a voice can be obtained from five sustained phonations. [Work supported by a Japanese Grant-in-Aid for Scientific Research.]

Matrix quantization for very low data rate voice coding

Matrix quantization extends the theory of vector quantization for voice coding to take advantage of the phonological and phonotactic properties of speech. The speech signal is first modeled by LPC spectra at 10‐ms time intervals. By adopting an acoustically defined (and demisyllable like) automatic segmentation procedure, consecutive LPC spectra are grouped into LPC coefficient matrices. A distortion measure that consists of nonlinear time warping and a standard LPC distance metric is also developed for comparing the “likeliness” of two matrices. The matrix codebook is generated from a speech database using a simple mini‐max procedure and the matrix distortion measure. A nearest‐neighbor full search procedure is used during speech encoding. Data rates under 100 bit/s can be achieved for the LPC spectra. Results for a single speaker experiment based on the matrix quantization method and subsequent research on very similar techniques will be discussed.

Erratum: Improved sequential linear prediction by selective time-domain coefficient extraction

For voiced speech, LPC coefficients obtained from sequential adaptation techniques are maximally perturbed roughly at the instant of glottal closure. Coefficient extraction at or immediately beyond this instant results in a poor vocal tract model estimate. A method is described to enhance the accuracy of the estimate by selecting coefficient vectors based on their time-domain convergence.

A construction of pole‐deviation tracking filter

AbstractThis paper describes a construction method for the pole‐deviation tracking filter, which is a computing system to follow the variation‐in‐time of the poles of the signal waveform. Until now, the methods such as filter‐bank, FFT and LPC, are used to track the deviation of poles, assuming the stationarity within the frame. There are some methods in nonstationary speech analysis that assume the variation within the frame. However, they have a problem in that the error may accumulate since the locations of poles are calculated after correcting the LPC coefficients. The method proposed in this paper, on the other hand, is an extension of the idea of POLPEC I, which we have proposed as a means of speech recognition. It assumes the variation within the frame, and the pole deviation is calculated directly from the waveform. By this scheme, the location of the pole is not averaged within the frame and the error does not accumulate. Neither is it necessary to solve an algebraic equation of higher order to determine the pole location. The result of simulation using synthesized waveform and the analysis of C‐V type speech also are discussed.

An improved LPC algorithm for voiced-speech synthesis

Current LPC synthesizers have a problem when the driving function is periodic; an output synthesized from LPC analysis has spectral properties which are not identical to those of the analyzed input. These distortions prevent a vocoder from being truly transparent for many voiced sounds. An iterative method, which modifies the conventional LPC coefficients such that the autocorrelation function of the output becomes identical to that of the input, is presented. A methodology for keeping the increased computational cost to a minimum is also to be discussed.

A comparison of two transforms from speech signals to tactile signals

Various methods of transforming speech signals into tactile signals have been investigated. Comparison of the results, however, is complicated by differences in training methods, training times, feedback, etc. In order to obtain a direct comparison between LPC-derived area function coding and spectral coding, naive subjects were trained successively on both coding schemes, using identical paradigms and speech tokens. The subjects were tested on identification and discrimination of four different b-V sounds presented to the index finger via a 6 × 10 portion of a vibrotactile array (Optacon). The vowels were /i/, /ε/, /ɑ/, and /u/, and three tokens of each b-V were used. The utterances were first processed to yield LPC coefficients, and then coded either in area function form or in spectral form using a technical mel vocoder. A “line” display was employed so that the number of active vibrators was constant for all stimuli. The durations of the stimuli were approximately equal. There were five sessions per week of approximately 35 min each, extending for up to eight weeks. The relative merits of the coding schemes, as well as the testing methods, are considered. [This work supported by NSF.]

The use of LPC coefficient interpolation to improve the naturalness of sentence formation by word concatenation

A recent experiment [Sanders et al., J. Acoust. Soc. Am. Suppl. 1 64, S1(A) (1978)] showed that an important contributor to the unnaturalness of sentence formation by word concatenation was the use of individually recorded words (isolation forms). Experience at Stanford has shown interpolation and quantization of LPC formant frequencies and bandwidths to be effective as methods of data rate reduction. Another use for interpolation is to improve the perceived naturalness of synthesized speech by smoothing the transitions between independently recorded sounds. Several considerations for successful word-boundary interpolation are: (1) syntactic, semantic and phonetic conditions, (2) the interpolation curve—linear or more complex, (3) the durations and locations of the interpolation region in individual words, and (4) the use of a specially prepared “demisyllabic bridge” to connect words. The relative importance of smoothing transitions involving high-frequency function words (articles, auxiliary verbs, etc.) is compared to interpolating all the junctions in a sentence. [Work supported by NSF grant SED-77-09698.]

Statistical properties of an LPC distance measure

Several distance measures have been proposed for comparing sets of LPC coefficients. The most popular one has been the "log likelihood ratio" proposed by Itakura [1]. In this paper we discuss this measure (strictly speaking, a somewhat generalized version of it) from both a theoretical and a practical point of view. We derive its statistical properties both when the reference vector is known and when it is estimated from the data. We also show how these properties are affected by windowing, additive noise, and preemphasis. We present results of extensive simulations in support of the theoretical predictions. Finally, we argue that de Souza's [2] recent criticism of this measure is unjustified.

A phrase recognizer using syllable-based acoustic measurements

A system for the recognition of spoken phrases is described. The recognizer assumes that the input utterance contains one of a known set of allowable phrases, which may be spoken within a longer carrier sentence. Analysis is performed on a syllable-by-syllable basis with only the strong syllables considered in the recognition process. Each strong syllable is represented in terms of a set of distinguishing acoustic measurements taken at time points in and around the syllable nucleus. Phrases are represented as sequences of strong syllables. All parameters used in recognition are derived from LPC coefficients. Input speech is limited to 3.3 kHz upper frequency. Recognition is completed within 1-3 s after the utterance is spoken. An interactive training facility allows flexible composition of key phrase sets. Testing was performed for a number of phrase sets each containing ten or fewer phrases, and included equal numbers of talkers used in training and talkers not used in training. Average phrase recognition accuracy was 95 percent when parameters were derived from unquantized (i.e., 16-bit) LPC coefficients and 90 percent when the LPC coefficients were transmitted to the recognizer across the ARPA network at 3500 bits/s. The recognizer has been incorporated into a user interface system where the parameters required to set up a point-to-point ARPANET voice connection can be established remotely by voice.

On creating reference templates for speaker independent recognition of isolated words

The three aspects of a statistical approach to a pattern recognition problem are the selection of features, choice of a measure of similarity, and a method for creating the reference templates (patterns) used in the statistical tests. This paper discusses a philosophy for creating reference templates for a speaker independent, isolated word recognition system. Although there remain many unanswered questions both about how to select appropriate features for recognition, and how to measure similarity between sets of features, such issues are not discussed here. Instead we concentrate on methods for creating the reference templates. In particular, a method of combining word patterns from a number of speakers is proposed in which a clustering type of analysis is used to determine which patterns are merged to create a word template. The creation of multiple templates, based on this method, is discussed and is shown to be of substantial value for as few as eight speakers in the training set. To test the ideas proposed here, a 54 word vocabulary word recognition system was implemented. All input words were recorded off a standard telephone line. The features used were the LPC coefficients of an 8-pole analysis, and the simple Itakura distance measure was used to measure similarity between patterns. With word templates obtained as described above, recognition accuracies of 85 percent were obtained in a forced choice recognition test on the 54 word vocabulary using eight new speakers. The correct word was within the top five choices 98 percent of the time. Using a strategy in which all the training words were used to create the templates, the recognition accuracy fell to 77 percent, and the correct word was within the top five choices only 89 percent of the time.

Statistical tests and distance measures for LPC coefficients

This paper considers the problem of comparing two sets of (LPC) coefficients or, more generally, that of comparing two short segments of speech via LPC techniques. It is shown that Itakura's prediction-residual ratio is intuitively unsatisfactory and theoretically misleading as a distance measure. Two slower, but more accurate statistical means of comparison are suggested, and these are supported by evidence from a simulation study.

Tandem Connections of Wideband and Narrowband Speech Communication Systems Part 2-Wideband-to-Narrowband Link

In this paper the tandem link of a 16 kb/s Continuously Variable Slope Delta modulator (cvsd) waveform coder and a 2.4 kb/s Linear Predictive Coding (lpc) vocoder is studied. Of prime concern are the effects of the cvsd coder on the lpc vocoder analyzer. In particular the problems involved in making a reliable voiced-unvoiced decision, estimating pitch period, and estimating lpc coefficients from the coder output are studied. It is shown that lpc coefficient estimation from the cvsd output is highly inaccurate. An analytical distortion measure (an lpc distance) is used to show the magnitude of the distortion introduced by the coder as a function of the signal gain into the cvsd coder. Although the remainder of the lpc analysis (i.e., pitch detection, voiced-unvoiced decision, and gain calculation) can be performed reasonably accurately, the magnitude of the distortions in estimating the lpc coefficients is sufficiently large to make the vocoded speech barely intelligible and of poor quality.

Speech‐quality testing of some variable‐frame‐rate (VFR) linear‐predictive (LPC) vocoders

VFR transmission of LPC vocoder coefficients is a technique developed to reduce the average transmission rate without appreciable loss of quality. The technique transmits parameters at a variable rate in accordance with the changing characteristics of the speech signal. In order to assess the effectiveness of VFR transmission, we performed an experiment to compare it with two other methods for reducing the bit rate: (a) reducing the number of poles, and (b) increasing the quantization step size of the LPC coefficients (log–area ratios). Thirty‐two stimulus sentences were prepared by passing four utterances (2 sentences × 2 speakers) through eight vocoder systems in a 2×2×2 factorial design; two values were assigned to each of the three parameters: average frame rate, number of poles, and quantization step size. Eight listeners made seven‐point category ratings of quality degradation. The results of the experiment show that, of the three methods studied, the VFR technique produced the highest quality at any given transmission rate (or, equivalently, yielded the lowest bit rate for a fixed level of speech quality).

Analysis of the performance of a tandem connection of a 16-Kbps CVSD coder to a 2.4-Kbps LPC vocoder

In this paper the tandem link of a 16-Kbps CVSD waveform coder and a 2.4-Kbps LPC vocoder are studied. Of prime concern are the effects of the CVSD coder on the LPC vocoder analyzer. In particular the problems involved in making a reliable voiced-unvoiced decision, estimating pitch period, and estimating LPC coefficients from the coder output are studied. It is shown that LPC coefficients estimation from the CVSD output is highly inaccurate. An analytical distortion measure (an LPC distance) is used to show the magnitude of the distortion introduced by the coder as a function of the signal gain into the CVSD coder. Although the remainder of the LPC analysis (i.e., pitch detection, voiced-unvoiced decision, and gain calculation) can be performed reasonably accurately, the magnitude of the distortions in estimating the LPC coefficients is sufficiently large to make the vocoded speech barely intelligible and of poor quality.

Evaluation of a Statistical Approach to Voiced-Unvoiced-Silence Analysis for Telephone-Quality Speech

Recently, a statistical-decision approach to the problem of voiced-unvoiced-silence detection of speech was proposed by Atal and Rabiner. This method was found to perform well on high-quality speech. However, the five speech parameters used in the analysis were not found to be as good for telephone-quality speech. Thus, an investigation was undertaken to determine a suitable set of parameters that would provide a reliable voiced-unvoiced-silence decision across a variety of standard telephone connections. A large number of parameters (70) were included in the investigation, including 12 lpc coefficients, 12 correlation coefficients, 12 parcor coefficients, 12 lpc partial error terms, etc. Many of the parameters were immediately eliminated because they provided almost no separability between the three decision classes. The remaining parameters were used in a knockout optimization to determine the five best parameters to use for a voiced-unvoiced-silence analysis. Various error weights were investigated to see what types of errors occurred and how they could be minimized. Finally, the use of the Itakura two-pole spectral normalization was investigated to see its effect on the error scores.

Evaluation of a voiced-unvoiced-silence analysis method for telephone-quality speech

Recently a statistical decision approach to the problem of voiced-unvoiced-silence detection of speech was proposed by Atal and Rabiner [IEEE Trans. on Acoust. Speech Signal Process., June (1976)]. This method was found to perform well on high-quality speech. However, the five speech parameters used in the analysis were not found to be as good for telephone-quality speech. Thus an investigation was undertaken to determine the five best parameters to use for a voiced-un-a reliable voiced-unvoiced-silence decision across a variety of standard telephone connections. A large number of parameters (69) were included in the investigation. including 12 LPC coefficients, 12 correlation coefficients, 12 PARCOR coefficients, 12 LPC partial-error terms, etc. Many of the parameters were immediately eliminated because they provided almost no separability between the three decision classes. The remaining parameters were used in a knock out optimization to determine the five best parametes to use for a voiced-unvoiced-silence analysis. Various error weights were investigated to see what types of errors occurred and how they could be minimized. Finally, the use of the Itakura two-pole normalization was investigated to see its affect on the error scores.

Triple-function voice coder (TRIVOC)

The first-formant region of the speech spectrum must be characterized with a high degree of accuracy in order to attain good speech quality and naturalness in a narrow-band speech processing system. This consideration led to the concept of using linear predictive coding for the first-formant region, vocoder channels for the upper spectrum, and detection and coding of the excitation function, to establish a narrow-band voice digitizer. This concept has advantages in the tradeoff of speech quality versus data rate, in comparison with the use of either channel vocoding or linear prediction alone. Such a triple-function voice coder (TRIVOC) has been implemented and investigated using a CSP-30 Digital Signal Processor. Processing algorithms have been established to permit comparison of various combinations of baseband and vocoder channels, LPC coefficients, and data rates ranging from 2400 to 4800 bits per second. Subjective results have confirmed the superior speech quality attainable by this method. Several versions of TRIVOC processing are presented, together with some results of subjective assessment of performance.