Automatic Speech Analysis Research Articles

Dynamique d’un fonctionnement psychologique à travers l’étude de l’énonciation

The aim of this study is to focus on the clinical aspects of four linguistic indicators that emerge from automatic speech analysis. From a theoretical point of view, the number of proposals deals with cognitive activity, the number of modelizations with emotional activity, the number of connections with judgment activity and the number of action verbs can refer to behavioral activity. To test these hypotheses we have studied two protocols of Thematic Aperception Test (TAT) randomized from non-clinical groups for the former and from a group of schizophrenic patients for the latter. The outcomes regarding the non-clinical protocols lead to the conclusion that the four indicators are coherent (α=0.93) and correlated; this confirms the clinical data on the cohesive personality of this young lady. The absence of correlations within the schizophrenic protocols reflects the dissociative syndrome of this patient. Finally, the study of the dynamics of this protocol, using our four linguistic criteria, confirms both the affective indifference and the increase of behavioral activity of this patient when the mentalization fails. After discussion and despite its capacity to describe clinical cases, the validity of this method needs to be completed (by relating the language indicators to production time stories) in further explorations, even if its reliability is good (α=0.93).

Comparative Study on Feature Selection and Fusion Schemes for Emotion Recognition from Speech

The automatic analysis of speech to detect affective states may improve the way users interact with electronic devices. However, the analysis only at the acoustic level could be not enough to determine the emotion of a user in a realistic scenario. In this paper we analyzed the spontaneous speech recordings of the FAU Aibo Corpus at the acoustic and linguistic levels to extract two sets of features. The acoustic set was reduced by a greedy procedure selecting the most relevant features to optimize the learning stage. We compared two versions of this greedy selection algorithm by performing the search of the relevant features forwards and backwards. We experimented with three classification approaches: Naive-Bayes, a support vector machine and a logistic model tree, and two fusion schemes: decision-level fusion, merging the hard-decisions of the acoustic and linguistic classifiers by means of a decision tree; and featurelevel fusion, concatenating both sets of features before the learning stage. Despite the low performance achieved by the linguistic data, a dramatic improvement was achieved after its combination with the acoustic information, improving the results achieved by this second modality on its own. The results achieved by the classifiers using the parameters merged at feature level outperformed the classification results of the decision-level fusion scheme, despite the simplicity of the scheme. Moreover, the extremely reduced set of acoustic features obtained by the greedy forward search selection algorithm improved the results provided by the full set.

Performance of an Automated, Remote System to Detect Vocal Fold Paralysis

The aim of this project was to evaluate the diagnostic accuracy of an automated, remote system for correctly identifying vocal fold paralysis. Consecutive patients presenting for vocal analysis at the Beaumont Hospital Voice Clinic were enrolled in this prospective, blinded study. Control patients were enlisted from routine otolaryngology clinics. All patients were assessed by standard history, clinical examination, and flexible laryngoscopy or videostroboscopy. The subjects were blindly assessed by remote voice analysis. Sustained phonation was recorded over a standard telephone network. Each recording was subjected to automated, remote analysis of extracted features, including measures of pitch perturbation, amplitude perturbation, and harmonics-to-noise ratio. The presence or absence of a vocal fold paralysis as determined by the automated classifier was recorded and correlated with clinical findings. Seventy-eight consecutive patients were enrolled in the study. The automated speech analysis system demonstrated 92% sensitivity and 75% specificity for detecting vocal fold paralysis. This pilot study, assessing an automated system that analyzes audiological data remotely over the standard telephone network, suggests that with further "training" it may become a reliable, simple, and convenient means for screening patients for voice disorders.

Telephony-Based Voice Pathology Assessment Using Automated Speech Analysis

A system for remotely detecting vocal fold pathologies using telephone-quality speech is presented. The system uses a linear classifier, processing measurements of pitch perturbation, amplitude perturbation and harmonic-to-noise ratio derived from digitized speech recordings. Voice recordings from the Disordered Voice Database Model 4337 system were used to develop and validate the system. Results show that while a sustained phonation, recorded in a controlled environment, can be classified as normal or pathologic with accuracy of 89.1%, telephone-quality speech can be classified as normal or pathologic with an accuracy of 74.2%, using the same scheme. Amplitude perturbation features prove most robust for telephone-quality speech. The pathologic recordings were then subcategorized into four groups, comprising normal, neuromuscular pathologic, physical pathologic and mixed (neuromuscular with physical) pathologic. A separate classifier was developed for classifying the normal group from each pathologic subcategory. Results show that neuromuscular disorders could be detected remotely with an accuracy of 87%, physical abnormalities with an accuracy of 78% and mixed pathology voice with an accuracy of 61%. This study highlights the real possibility for remote detection and diagnosis of voice pathology.

Cepstral analysis with nonuniform spectral weighting for spectral envelope extraction

AbstractThis paper proposes a method for extraction of the spectral envelope by a spectral evaluation criterion with nonuniform weighting on a log magnitude scale. In the traditional cepstral method, the spectral envelope is obtained by linear smoothing on the log scale. Consequently, the obtained log magnitude spectrum is affected by the low‐level portion of the spectral fine structure, one of the factors that has degraded the quality of the synthesized speech and the speech recognition rate. to avoid such a situation, the proposed method weights the peaks of the fine spectral structure more heavily than the valleys in the error criterion. Consequently, the resultant spectral envelope fits the peaks, i.e., the sampling points of the true spectral envelope for a voiced sound. This method does not require pitch extraction. In addition, the spectral envelope can be obtained stably both for voiced and unvoiced sounds. Thus, it is suitable for the automatic analysis of speech. the effectiveness of the method is demonstrated by some examples, and its relation to other methods, especially to the improved cepstral method, is discussed.

Should ASR front-end be insensitive to fundamental frequency? (perceptual shift of formant position due to fine harmonic structure of voiced speech

Due to the fine harmonic structure of the source spectrum of voiced speech, most automatic speech analysis techniques estimate formant positions that systematically deviate from the resonant frequencies of the vocal tract. This property is in conflict with linear source-filter models of speech production and an effort is made to minimize the effect of the fundamental frequency F0 on the analysis result [e.g., Hermansky et al., Proceedings ICASSP-83 (IEEE, New York, 1983), pp. 778–780]. Our experiments show that the human auditory perception has properties similar to the above described properties of automatic analyses. The method of confusion scaling is applied in paired comparison of synthetic vowel-like stimuli with varying F0 and the second formant frequency F2. The results show that the perceptual estimate of formant position goes through a cycle of positive and negative deviations as the relative F2/F0 frequency goes from one integral number to the next. The position is estimated the most accurately when the formant peak falls either into a harmonic peak or if it lies right between two harmonic peaks. The maximal deviation of the estimate due to harmonics of the F0 from the 120-Hz neighborhood in the 2-kHz second formant region is about ± 0.5%. Our findings imply that when modeling the speech perception, as in the automatic speech recognition, a certain degree of sensitivity of the analysis technique to F0 might be beneficial.

Automatic analysis of speech using parallel cellular pipelined processor

Early results of analyzing natural speech using a parallel processor were reported before this society in Minneapolis in the fall of 1984. This paper is a presentation of subsquent results obtained. The CYTO‐HSS™ parallel pipelined processor, used for analyzing speech spectrograms, is a work station for image analysis, optimized for cellular automata algorithms. The HSS has several advantages over typical Von Neumann type machines. One advantage is speed. In performing similar image processing tasks, the CYTO‐HSS™ is from 100 to 10 000 times faster than the VAX. The distinctive features of speech are described in mathematical morphological terms and the parallel processor detects the presence of the features. Examples of stop and fricative detection will be described and shown. Potential applications in speech analysis, automatic speech recognition, speech synthesis, bandwidth compression, and speech enhancement are discussed.

Automatic Speech Analysis and Recognition edited by Jean‐Paul Haton

First Page

Automatic speech analysis and recognition: Proceedings of the NATO advanced study institute held at Bonas, France

Automatic speech analysis and recognition: Proceedings of the NATO advanced study institute held at Bonas, France

The use of tracking rules in automatic formant analysis of speech

The J.S.R.U. computer‐simulated automatic formant analyser uses analysis by synthesis to achieve high reliability and accuracy in the matching of synthetic formants to the peaks of pitch synchronous short term spectra of natural speech [P.M. Seeviour, J. N. Holmes and M. W. Judd, IEEE Int. Conf. Acoust., Speech, Signal Processing, Philadelphia, PA, pp. 609–693 (1976)]. However, analysis by synthesis does not of itself necessarily provide consistent labeling of the peaks of the natural spectrum. The specific case of this problem as it appears in the J.S.R.U. formant analyser will be discussed. The task is to choose, for each 10 ms frame of speech, from among a number of attempts to produce a synthetic match to the natural spectrum. The restrictions on the algorithm, particularly the need to avoid excessive smoothing of formant movement, will be mentioned. Simple heuristic rules currently in use will be described and the performance obtainable with such rules will be illustrated by means of spectrograms and audio tape recordings.

Linear prediction analysis of speech based on a pole-zero representation.

Speech analysis and synthesis by linear prediction is based on the assumption that the short-time spectral envelope of speech can be represented by a number of poles. An all-pole representation does not provide an accurate description of speech spectra, particularly for nasals and nasalized sounds. This paper presents a method for characterizing speech in terms of the parameters of a pole-zero model. In this method, an impulse response representing the composite filtering action of the glottal wave, the vocal tract, the radiation, and the speech recording system is first constructed from the speech signal. This impulse response is obtained by performing several stages of all-pole LPC analysis. The pole-zero parameters are determined from the impulse response by solving a set of simultaneous linear equations. The method, being noniterative, is very suitable for automatic analysis of speech. The method has been applied to real speech data and the results show that the speech spectra derived from the pole-zero model agree very closely with the actual spectra derived by direct Fourier analysis.

Large speech data bases

It is easy enough to digitize a large amount of speech data. But before it can be effectively used in speech research, it must be cataloged to indicate position and descriptions of words and phones that are present in the data. In the absence of adequate tools, experts must do these tasks manually. Given interactive, automatic, and semiautomatic speech analysis programs, one can significantly improve the quality of the data base and the productivity of the experts. This paper describes the structure and characteristics of programs developed at Carnegie-Mellon University which interactively bootstrap themselves to generate symbolic descriptions of a given set of data. The present system contains programs for (1) generating environment and speaker adapted phone templates (2) interactive generation of a phone lexicon containing alternating pronunciation of words generated from data, and (3) a program for machine aided labeling of a given phrase or sentence giving the beginning and ending of each phone and each word. Using these programs, over three hours of connected speech data has been digitized, and analyzed, to generate symbolic descriptions of the data in terms of phones and words. Retrieval programs are then used to retrieve all symbols satisfying a given property.

Automatic segmentation of speech into syllabic units.

As a first step toward automatic phonetic analysis of speech, one desires to segment the signal into syllable-sized units. Experiments were conducted in automatic segmentation techniques for continuous, reading-rate speech to derive such units. A new segmentation algorithm is described that allows assessment of the significance of a loudness minimum to be a potential syllabic boundary from the difference between the convex hull of the loudness function and the loudness function itself. Tested on roughly 400 syllables of continuous text, the algorithm results in 6.9% syllables missed and 2.6% extra syllables relative to a nominal, slow-speech syllable count. It is suggested that inclusion of alternative fluent-form syllabifications for multisyllabic words and the use of phonological rules for predicting syllabic contractions can further improve agreement between predicted and experimental syllable counts. Subject Classification: 70.40, 70.60.

Scheme for Automatic Formant Analysis of Speech

The concept of a formant as representing a peak in the pressure spectrum is assumed to be applicable to both voiced and unvoiced speech. A scheme has been developed that estimates five formant frequencies and amplitudes. The cepstrum technique is used, along with intensity, zero-crossing, and slope-change information, to make voiced-unvoiced decisions and to estimate the fundamental voicing frequency. An attempt is made to detect bursts of energy (due primarily to stop consonants) in the time waveform and to analyze them with sufficient time resolution so that the burst characteristic is preserved. The formant estimating procedure is based on assumed formant exclusive domains in frequency space. Several smoothing procedures are used to remove discontinuities from the formant and fundamental frequency data, and the smoothed values are used to control a five-pole parallel synthesizer. The synthesizer is excited with a pulse train, noise, or a mixture of the two. Examples of the natural and synthetic speech are presented.