Quality Of Synthesized Speech Research Articles

Speaker-dependent-feature extraction, recognition and processing techniques

This paper discusses recent advances in and perspectives of research on speaker-dependent-feature extraction from speech waves, automatic speaker identification and verification, speaker adaptation in speech recognition, and voice conversion techniques. Speaker-dependent information exists both in the spectral envelope and in the supra-segmental features of speech. This individual information can be further classified into temporal and dynamic features. Speaker identification/verification methods can be divided into text-dependent and tect-independent methods. Although text-dependent speaker verification techniques have almost reached the level suitable for practical implementation, text-independent techniques are still in the fundamental research stage. Both supervised and unsupervised speaker adaptation algorithms for speech recognition have recently been proposed, and remarkable progress has been achieved in this field. Improving synthesized speech quality by adding natural characteristics of voice individuality, and converting synthesized voice individuality from one speaker to another, are as yet little exploited research fields to be studied in the near future. Research on speaker-dependent information is one of the most important future directions for achieving advanced speech information processing systems.

An efficient algorithm for articulatory synthesis.

One of the methods for efficient computation of articulatory synthesis is to discretize the area function into a small number of sections such as 10 or 12 resulting in a poor spatial resolution. The articulatory data have to be represented by 30 or more sections for a good spatial resolution requiring a large number of computations. However, good quality speech synthesis has been obtained using formant or LPC-based methods with filters of 10 or 12 coefficients for voiced sounds and 4 or 6 coefficients for unvoiced sounds. It is proposed that a computationally efficient method be used to obtain high quality synthesis using an articulatory model with a large number of sections (30 or more). It can be shown that when the adjacent sections are of equal area, the computations can be pruned without any loss of accuracy. For a given computational load, using an effective quantization rather than discretization of area function, one can achieve much lower spectral error in the transfer function. Trade-off relationships between spectral error and computational load as a function of quantization and discretization are reported.

Excitation waveform extraction for pitch control in residual‐excited LPC speech synthesis

The residual‐excited LPC is one of the most effective techniques for producing high‐quality synthetic speech. However, this technique has difficulty controlling pitch frequency when applied to realizing synthesized speech with a pitch contour different from original speech or when applied to synthesis of arbitrary speech created with concatenated spoken units. In previous methods, pitch‐period LPC residual waveforms that extracted pitch synchronously have been used to control pitch frequency. However, the extraction position (Eρ) and window length (Ew) are very critical to synthesized speech quality, and cause voice quality deterioration. This paper proposes a new method of excitation waveform extraction that automatically determines Eρ and Ew using spectral envelope distortion criteria between input and synthetic speech. Subjective evaluation experiments indicate that the pitch frequency pattern can be changed with a relatively small deterioration in quality. Application of this method to arbitrary speech synthesis will also be discussed.

Role of prosody in cognitive process of spoken language

AbstractThe authors are developing an arbitrary text speech synthesis system, and aiming at the improvement of the synthesized speech quality, the effect of the prosodic information data of the speech, e.g., pitch structure and the time structure, on the speech intelligibility was evaluated by a listening experiment using synthesized speech. It was verified quantitatively by the result of experiment that the pitch information works efficiently in the speech listening without attention concentration, aiding the understanding of the speech. It was seen that the accent information is especially important, so that the accent should be attached to the text speech synthesis system resulting in an accuracy as high as 98 percent or more. It was shown that the existence or nonexistence of the phrase component does not affect speech intelligibility. Some results of the experiment suggest a direction toward improving the performance of the speech recognition system. It was seen that if the phrase recognition rate of 70 percent is achieved in the speech recognition system from the acoustic viewpoint, the rate can be improved up to 90 percent or more without using the pitch information by employing the lower‐ and higher‐level linguistic knowledge' processing. The importance of these factors in improving the intelligibility are in the order of lexicon, context, and accent information.

Vector quantization of speech signals using principal component analysis

AbstractA new method of designing a vector quantizer is presented for bandwidth compression of speech signals, and some experimental results are shown. In this method, spectral parameters are extracted first from the DFT spectrum of input speech signals using a psychological frequency scale‐the so‐called Mel scale. This parameter is called the Mel‐scaled spectrum. The number of Mel‐scaled spectrum is reduced and the cepstrum of this parameter is calculated. Then this Mel‐scaled cepstrum is vector quantized and the codebook‐vector of the vector quantizer is determined by the algorithm using principal component analysis. With this algorithm, codebook‐vectors can be designed considering the statistical characteristics of the Mel‐scaled cepstrum. Also, the reduction of parameters by Mel‐scale can decrease the size of the codebook memory without greatly degrading the synthesized speech quality.Using the forementioned method two codebooks are designed: one contains 256 vectors, and the other contains 2048 vectors. The quantization error is compared with those designed by the well‐known LBG algorithm. The simulation results show that the codebooks designed by the proposed method present less quantization error and degradation of synthesized speech quality than those designed by LBG algorithm.

Multipulse linear predictive synthesis in text‐to‐speech systems

Multipulse linear prediction is a coding technique that can provide extremely high quality speech synthesis [B. S. Atal and J. R. Remde, IEEE Proc. ICASSP82, 614–617 (1982)]. It is therefore of interest to examine whether the technique can be used to provide correspondingly good quality in a text‐to‐speech system. In such a system a prosodic contour is imposed on a set of concentrated speech units. Various speech units (e.g., words and diphones) have been tried. First results suggest that the word is the appropriate unit to be used in such a system. The techniques for pitch and timing alteration, speech unit concatention, and the effects on the resulting synthetic speech will be discussed and demonstrated.

Human Factors and Synthetic Speech

Recent advances in linguistics, speech science, psychology, and especially in computers have made unlimited text-to-speech conversion systems a practical reality. However, the use of audio output from a computer poses special problems in ergonomics, most of which have not been dealt with in the literature. To provide high quality speech synthesis in a useful and usable system requires research in several areas (prototype applications, general design issues for touch-tone phone/speech interfaces, and improving the speech signal itself). In this paper, we review recent work in our laboratory in each of these areas. We close with a discussion of the implications of our work to date for the development and use of speech synthesis.

A proposed adaptive control method for packet transmission delay in packet speech system

AbstractThere are two basic problems in speech packet switching systems. One is how to guarantee the synthesized speech quality against fluctuation of the mean and variance of packet transmission delay in the network due to traffic fluctuation in packet switching. the other is the problem of deciding the beginning time of the talkspurt synthesis (i.e., when to begin talkspurt synthesis at the receiver side). Under the assumption that the packet transmission delay time is approximated by an exponential function, it is shown that the jitter absorbing time for each talkspurt can be calculated by observation of the packet arrival time at the receiver side. Using this result, an adaptive control method is presented to cope with the above two problems, where the jitter absorbing time is changed according to network traffic changes and the talkspurt synthesis time is determined from several packet arrival times. It is shown from simulation and numerical calculation that application of this control method to a speech packet switching system permits guaranteeing the voice quality in spite of traffic fluctuation in the network and reducing the probability of late beginning of talkspurt synthesis.

The sensitivity of LPC synthesized speech quality to the imposition of artificial pitch, duration, loudness, and spectral contours

In attempting to achieve more natural sounding synthesized speech, it is important to determine how speech quality is sensitive to the manipulation of spectral information and prosodic components: pitch, duration (rhythm), and loudness, in synthesizing sentences. To assess the impact of each of these on the perceived naturalness of synthesized speech, we have modified a set of recorded and LPC‐analyzed sentences, so that one or more of these components is rule‐generated from information available from a lexicon of individually recorded and LPC‐analyzed words. Substitution of lexical spectral information for the observed sentential spectra, or imposition of a syntactically generated pitch contour [A. Levine, J. Acoust. Soc. Am. 60, S76(A) (1976)], are two of the modifications we will be testing here. Results will be reported from an experiment in which listeners judged the naturalness of these modified sentences. The usefulness of evaluating modified natural speech as a method for studying the quality of synthesized speech will also be discussed.

Laryngeal Activity and the Analysis and Synthesis of Speech

The activity of the larynx is a function of the tensions of the laryngeal muscular system, the mass of those parts of the vocal cords that are in motion, the losses caused by the collision of the vocal cord's margins, and the aerodynamic forces that act upon the vocal cords. The larynx is coupled through these aerodynamic forces to the rest of the respiratory system. The subglottal air pressure is affected by the resonances of the lungs and trachea. When the fundamental frequency of the glottal waveform or the lower harmonics of the fundamental frequency excite these resonances, the peak subglottal air pressure builds up. The superglottal air pressure is a function of the air flow through the vocal tract. Pressure transients result when constrictions occur during voicing. Variations in the glottal waveform, therefore, result from period to period, which causes rapid variations in the glottal source's spectrum and periodicity. This paper discusses the results of analyses of the acoustic waveform and high-speed ,notion pictures of the larynx and the relevence of these experimental results to the preservation of natural quality in speech synthesis.