Speech Processing Systems Research Articles

A PC-KIMMO-based Bi-directional Graphemic/PhoneticConverter for Modern Greek

This report confronts the problem of automatic conversion from graphemic to phonetic transcription and vice versa for the Modern Greek language. A single representation is used for both directions of word transformation, based on PC-KIMMO, a development environment originally used for the implementation of two-level morphological processors. The fifty-two two-level rules that are presented have been tested on a set of the 10,000 most frequent Modern Greek words and they describe the Greek phonology in adequate detail for use by a speech-processing system.

Method for processing speech signal in speech processing system

A method for processing an input speech signal to be applied to a CELP vocoder has the steps of obtaining preliminary pitch search intervals by a preprocessing autocorrelation expression from a pitch lag of a synthesized speech signal which is synthesized from a residual signal of the input speech signal; computing coefficients of pitch filter with respect to the preliminary pitches; searching a high interval in the autocorrelation; and removing the remaining interval other than the high interval in the pitch lag. Since the present invention proposes a speech processing method which uses only a high interval in autocorrelation of a voice waveform in pitch-searching, and where such a speech processing method is embodied in a CELP vocoder, total computation time of the CELP vocoder can be decreased 37% or more without lowering speech quality. Therefore, a digital signal processor, which is low in price and is slow in speed, can be embodied in a CELP vocoder.

CS-ACELP speech compression system with adaptive pitch prediction filter gain based on a measure of periodicity

A speech coding system employing an adaptive codebook model of periodicity is augmented with a pitch-predictive filter (PPF). This PPF has a delay equal to the integer component of the pitch-period and a gain which is adaptive based on a measure of periodicity of the speech signal. In accordance with an embodiment of the present invention, speech processing systems which include a first portion comprising an adaptive codebook and corresponding adaptive codebook amplifier and a second portion comprising a fixed codebook coupled to a pitch filter, are adapted to delay the adaptive codebook gain; determine the pitch filter gain based on the delayed adaptive codebook gain, and amplify samples of a signal in the pitch filter based on said determined pitch filter gain. The adaptive codebook gain is delayed for one subframe. The pitch filter gain equals the delayed. adaptive codebook gain, except when the adaptive codebook gain is either less than 0.2 or greater than 0.8., in which cases the pitch filter gain is set equal to 0.2 or 0.8, respectively.

Speech processing system and method for enhancing a speech signal in a noisy environment

A speech processing system (30) operates in a noisy environment (20) by performing adaptive prediction between inputs from two sensors positioned to transduce speech from a speaker, such as an accelerometer and a microphone. An adaptive filter (37) such as a finite impulse response (FIR) filter receives a digital accelerometer input signal, adjusts filter coefficients according to an estimation error signal, and provides an enhanced speech signal as an output. The estimation error signal is a difference between a digital microphone input signal and the enhanced speech signal. In one embodiment, the adaptive filter (37) selects a maximum one of a first predicted speech signal based on a relatively-large smoothing parameter and a second predicted speech signal based on a relatively-small smoothing parameter, with which to normalize a predicted signal power. The predicted signal power is then used to adapt the filter coefficients.

Multi-pulse analysis speech processing system and method

A speech processing system and method are disclosed. In one embodiment of the present invention, the system includes at least a maximum likelihood quantization (MLQ) multi-pulse analysis unit (item 21) operating on a target vector (item 26). The MLQ multi-pulse analysis unit typically determines an initial gain level for the multi-pulse sequence and performs single gain multi-pulse analysis (MPA) a number of times, each with a different gain level. The pulse sequence which most closely represents the target vector is provided as an output signal (item 38). In another embodiment, the system includes at least a pulse train multi-pulse analysis unit wherein the target vector is modeled as a series of pulse trains. Each pulse train comprises a plurality of single gain pulses, wherein each pulse is at a position which is a pitch value distance apart from the previous pulse in the pulse train. Combinations of maximum likelihood analysis with pulse trains are also part of the present invention.

Synergetics and 'insight' strategy for speech processing

The methodology and developmental technology were created for a speech understanding system based on synergetics and semantic-pragmatics ideas. A real speech processing system GOLOS-2 was then designed as a modular assembly whose modules correspond to similar levels of man's verbal and mental processes. Its behaviour is regulated by a through understanding 'insight' procedure that functions as a computer analogue for a pragmatic-communicative operator of human verbal/mental activity

Hearing with bionic ears [cochlear implant devices

According to speech perception-rate data continuous interleaved sampling (CIS) and spectral maxima sound processor (SMSP) techniques are, and probably will be, the best speech processing strategies for multichannel electrode cochlear implant devices. From packaging and power-consumption viewpoints, today's speech processing systems are very big and are, therefore, worn on the body and consume large electric power. Next-generation cochlear implant devices would be more compact, low-power products that would be worn behind or in the ear. It is clear that mixed-signal, high-density, and low-power design techniques are required to satisfy compactness, as well as low-power consumption features to realize intelligent speech sensation for the implantees. The especially critical design consideration of power supply lifetime and efficiency might be increased by using new promising technologies like microelectromechanical systems (MEMS).

Using Neural Networks to Determine Sugeno Measures by Statistics

To replace the traditional weighted average method, Choquet integrals or Sugeno integrals with respect to fuzzy measures are used to obtain a synthetic evaluation of a given object (or its quality, function, etc. respectively) with multi-attribute. Generally, it is not easy to determine fuzzy measures in real problems due to the subjectivity of human thinking. It is even much more difficult than determining weights in the weighted average method, because of the nonadditivity of fuzzy measures. This paper uses a neural network algorithm to optimize the inverse problem of synthetic evaluation, and thus to determine Sugeno measures by the Choquet integral and statistics of given data. Since the Choquet integral is a generalization of the weighted average method, this technology has a broad applicability in areas of multivariate analysis, decision making, pattern recognition, image and speech processing and expert systems. Copyright © 1996 Elsevier Science Ltd.

Generating stressed speech from neutral speech using a modified CELP vocoder

The problem of speech modeling for generating stressed speech using a source generator framework is addressed in this paper. In general, stress in this context refers to emotional or task induced speaking conditions. Throughout this particular study, the focus will be limited to speech under angry, loud and Lombard effect (i.e., speech produced in noise) speaking conditions. Source generator theory was originally developed for equalization of speech under stress for robust recognition (Hansen, 1993, 1994). It was later used for simulated stressed training token generation for improved recognition (Bou-Ghazale, 1993; Bou-Ghazale and Hansen, 1994). The objective here is to generate stressed perturbed speech from neutral speech using a source generator framework previously employed for stressed speech recognition. The approach is based on (i) developing a mathematical model that provides a means for representing the change in speech production under stressed conditions for perturbation, and (ii) employing this framework in an isolated word speech processing system to produce emotional/stressed perturbed speech from neutral speech. A stress perturbation algorithm is formulated based on a CELP (code-excited linear prediction) speech synthesis structure. The algorithm is evaluated using four different speech feature perturbation sets. The stressed speech parameter evaluations from this study revealed that pitch is capable of reflecting the emotional state of the speaker, while formant information alone is not as good a correlate of stress. However, the combination of formant location, pitch and gain information proved to be the most reliable indicator of emotional stress under a CELP speech model. Results from formal listener evaluations of the generated stressed speech show successful classification rates of 87% for angry speech, 75% for Lombard effect speech and 92% for loud speech.

Speech synthesis using a nonlinear modeling of vocal folds vibration effect

From a theoretical viewpoint, the vocal folds vibration affects the vocal tract transfer characteristics through nonlinear time-varying interaction between the glottis and vocal tract. However, ordinary speech processing systems have neglected those effects under the assumption of linearity of production systems. But if one tries to improve the voice quality to be more natural or human-like, it becomes crucial to investigate and model such effects. For this purpose, analytic exeriments were conducted on the vocal folds vibration effects [Ohmura et al., ICSLP96]. Based on these experimental results, two types of speech synthesis models are proposed: one is a wave function model in which formant energy dumping is given by a time window function. The other is a second-order nonlinear differential equation in which the formant energy dumping pattern is controlled by its friction term. Its performance was confirmed by a preference test on speech wave reconstruction. A speech sample set used in this test contains 60 sentences uttered by four speakers. Evaluators listened to two kinds of reconstructed speech samples: one is synthesized by the proposed method and the other is synthesized by an ordinary LPC-based method. Results show that the proposed method is clearly superior in its quality.

Robust speaker recognition: a feature-based approach

The future commercialization of speaker- and speech-recognition technology is impeded by the large degradation in system performance due to environmental differences between training and testing conditions. This is known as the "mismatched condition." Studies have shown [l] that most contemporary systems achieve good recognition performance if the conditions during training are similar to those during operation (matched conditions). Frequently, mismatched conditions axe present in which the performance is dramatically degraded as compared to the ideal matched conditions. A common example of this mismatch is when training is done on clean speech and testing is performed on noise- or channel-corrupted speech. Robust speech techniques [2] attempt to maintain the performance of a speech processing system under such diverse conditions of operation. This article presents an overview of current speaker-recognition systems and the problems encountered in operation, and it focuses on the front-end feature extraction process of robust speech techniques as a method of improvement. Linear predictive (LP) analysis, the first step of feature extraction, is discussed, and various robust cepstral features derived from LP coefficients are described. Also described is the afJine transform, which is a feature transformation approach that integrates mismatch to simultaneously combat both channel and noise distortion.

Role of articulation in speech perception: clues from production.

The relative roles played by signal properties and nonsignal information in speech perception are first examined. The evidence strongly suggests that phonetic percepts are never knowledge-innocent records of the raw signal. That conclusion is drawn not only about "higher" levels of language processing, but is seen to apply also to the perception of elementary phonetic stimuli. A review of a broad range of facts about production highlights the fact that speech production is adaptively organized. That circumstance suggests that the signal does not encode articulatory or acoustic/auditory invariants, but plays the role of supplementing the multimodal information already in place in the listener's speech processing system. It is accordingly proposed that phonetic signals are not invariants wrapped in "noise," but are products of listener-dependent adaptations that transform speech patterns in principled and, therefore, interpretable ways. Do listeners form speech percepts by way of intermediate articulatory representations? There seem to be strong both theoretical and methodological reasons to doubt that they do.

Interactive voice technology development for telecommunications applications

This paper describes the essential speech processing techniques for interactive voice applications in the telecommunications field. These techniques include speech recognition and speech synthesis, both of which aim to make interactive speech communications between man and machine more natural. Keyword spotting, background noise effects reduction, and speakers and/or telephone adaptation techniques are considered essential in speech recognition in order to allow a more natural voice input as well as an adequate robustness against environmental variabilities. In the area of text-to-speech synthesis, we propose a rule-based synthesis method applicable to the Japanese language, aiming to produce high quality speech. The commercial system ANSER of a former project is also described as an example of an interactive speech processing system. Finally, a recently developed speech recognition server which includes a vocabulary-flexible recognition function is described. It is meant to illustrate the concept of the techniques it employs which allow its range of applications to be easily extended and also allow it to adapt itselt to the changes which are rapidly occurring in the telecommunications field.

Visible speech improves human language understanding: Implications for speech processing systems

Evidence from the study of human language understanding is presented suggesting that our ability to perceive visible speech can greatly influence our ability to understand and remember spoken language. A view of the speaker’s face can greatly aid in the perception of ambiguous or noisy speech and can aid cognitive processing of speech leading to better understanding and recall. Some of these effects have been replicated using computer synthesized visual and auditory speech. Thus, it appears that when giving an interface a voice, it may be best to give it a face too.

Explicit estimation of speech boundaries

Estimation of signal length is one of the first steps in many speech processing systems, for it provides us with the initial and final points between which the speech signal is comprised. Nevertheless, endpoint detection can be affected by different types of background noise over which we have no control. This study presents an explicit detector that aims to resolve the problems created by two types of noise: quasi-stationary background noise and the noise generated by the speaker. The system is based on very simple measurements, energy, zero crossing rate and band crossings, and has been implemented on a personal computer. We have tested the system under real-life conditions, and it has attained a success rate of 97-98% in the location of the extremes for both the left and the right limit of a series of words comprising all the possible beginnings and endings in Spanish.

Role of articulation in speech perception: Clues from production studies

The relative roles that signal properties and nonsignal information play in speech perception are examined first. The evidence strongly suggests that ‘‘the input is never into a quiescent or static system.’’ Thus phonetic percepts are never knowledge-innocent records of the raw signal. Drawing this conclusion not only with reference to ‘‘higher’’ levels of language processing, but also for the perception of elementary phonetic stimuli, a broad range of facts about production are considered. Is it organized to permit listeners to recover articulations? Or is it shaped by a demand for (relative or absolute) acoustic stability despite articulatory imprecisions? Is it controlled to generate auditory invariants? The search for answers highlights the adaptive nature of speech production, which suggests that the signal does not encode articulatory, acoustic, or auditory invariants but plays the role of supplementing the multi-modal information already in place in the listener’s speech processing system. Phonetic signals are not invariants wrapped in ‘‘noise’’ but are products of listener-dependent adaptations that transform speech patterns in principled, and therefore interpretable, ways. Do listeners recover articulations? The present review leads one to doubt that they do.

Computational aids for the study of prosody

In the past few years, corpus-based methods of inquiry have yielded significant insights into the structure and role of prosody in human speech. Efforts are currently underway to discover the relationship between prosody and syntax and discourse, and to develop automated speech processing systems (both for synthesis and recognition) that take advantage of the information contained in prosody. These efforts, however, are critically dependent upon the availability of large speech corpora in which the relevant prosodic phenomena have been consistently transcribed. If the development of such a corpus is to be cost effective or, indeed, if prosodic cues are to be detected in automated systems, computational tools that facilitate and, where possible, automate the transcription process must be made available. In this paper, some the tools currently available will be presented and their performance and utility reviewed. In particular, a new tool for generating accurate, time-aligned phonetic transcriptions of spontaneous speech will be discussed, and quantitative measures of its performance will be presented. Integration of this tool into a comprehensive prosodic labeling environment will also be discussed.

Stability analysis on the time-domain finite-difference solution of a basilar-membrane vibration model: Application to acoustic signal processing

In the past several years, models for nonlinear wave motions in the basilar membrane based on the hydrodynamic principle and on the biophysical mechanisms of the cochlear function have been developed. The significance of such models can be projected to be their potential applications in designing front-end components of advanced speech processing systems, as well as in serving a powerful tool for studying information processing in central parts of the auditory system. Unfortunately, computer simulations of the models in the past had been exceedingly slow and sometimes unstable, which had prevented use of the models in many possible speech-processing applications. In this paper numerical properties of a finite-difference scheme for the time-domain solution of a one-dimensional nonlinear transmission line basilar membrane model are studied. In particular, the von Neumann method is applied to obtain a sufficient condition under which the finite-difference solution to the model is guaranteed to be stable. The utility of this condition is that it determines the optimal time and spatial mesh sizes, thereby guaranteeing a minimal amount of computation in obtaining stable and accurate model outputs. Also presented are preliminary results on applying the basilar-membrane model, based on the most efficient computation schemes derived from the stability analysis, to process speech signals.

Discharge pattern in the auditory nerve evoked by vowel stimuli: A comparison between acoustical and electrical stimulation

Single channel cochlear implants only transmit the time structure of the electrically coded input signal. All nerve fibres show similar thresholds for monopolar round window stimulation, i.e., activation does not depend on their site of origin. To investigate the fine structure of the firing pattern elicited by stimulation with an analogue coded speech processing system (VIENNA 1-channel implant), cats were electrically stimulated with German steady-state vowels at the round window. Single fibre activity was recorded from primary auditory fibres and period histograms were calculated. The electrically evoked impulse patterns were compared with those from acoustic stimulation with the same vowels. With acoustic stimulation, the response of a fibre depends on the individual characteristic frequency (CF) with regard to the fundamental F0 and the formants F1, F2 and F3 of the vowels, the spontaneous activity of the fibre and the sound level. The evoked firing pattern was used to calculate period histograms, the frequency content of which was analysed by Fourier transformation. With electrical stimulation in the threshold range, an action potential is strongly synchronized to a cathodic peak of the current within one period of F0. With increasing current level 3–5 impulses can be locked to the same period. The timing of the short intervals is determined by the relative refractory period and current peaks (negative or positive) caused by the dominant higher formant F2 or F3. The acoustically evoked patterns are specific for the CF of the neuron and represent the spectral information of the different vowels. The electrically evoked patterns are uniform and only small differences in the short intervals between successive nerve pulses with suprathreshold stimulation may enable the implanted patient to differentiate vowels.

Speech processing system

First Page