Hans Werner Strube

Linear prediction on a warped frequency scale

Linear prediction is considered with respect to a nonlinear frequency scale obtained by a first‐order all‐pass transformation. The predictor can be computed from a frequency‐warped autocorrelation function obtained from the power spectrum or by a direct linear transformation of the original acf. Three numerical procedures are compared. Alternatively, the predictor can be determined from a covariance matrix or (adaptively) from continuously formed correlations, suitably defined according to the all‐pass transformation. Prediction‐error minimization and spectral flattening are no longer equivalent criteria. In the synthesis part of a vocoder or APC system, no inverse transformation is required, since the direct form of the analysis and synthesis filters can be modified so as to immediately realize the warped transfer function. Single‐word intelligibility is compared for a predictive vocoder on a ’’Bark’’ scale and a linear frequency scale. The Bark scale yields results around 90% even at predictor orders of...

Determination of the instant of glottal closure from the speech wave

For vowels excited by vigorous glottal vibrations, the instant of glottal closure is tentatively identified with the moment of strongest excitation (at not too low frequencies) and worst linear predictability. Some predictor methods for its determination are reviewed, which do not always yield reliable and unequivocal results. Then Sobakins method using the determinant of the autocovariance matrix is examined critically and reinterpreted such that the determinant is maximum if the beginning of the interval on which the autocovariance matrix is calculated coincides with the glottal closure. This hypothesis is tested by comparison with the predictor methods and by looking at the inversely filtered waveforms and the formants obtained from predictors determined on a shifted interval. The determinant method seems to be very reliable even for otherwise difficult cases, such as the vowel /u/.

Speech enhancement by nonlinear multiband envelope filtering

Starting from a known relation between modulation transfer function (of a room) and speech intelligibility, we tried to enhance speech corrupted by reverberation or noise by suitably filtering the envelope (power) signals in critical frequency bands. Two methods based on FFT-overlap-adding and onlinear prediction, respectively, have been developed. Linear envelope filtering was not successful for neither pre- nor postprocessing, even though the objective SNR was improved. However, inserting a logarithmic nonlinearity or, better, a logarithmic one becoming linear at high levels, did raise intelligibility in the presence of noise when the filter effected strong compression at very low and no change or expansion at higher modulation frequencies. Since noise masking can serve as a model for recruitment in sensorineural hearing impairment, the method might be used in future hearing aids.

A quasiarticulatory speech synthesizer for German language running in real time

A system for simple quasiarticulatory speech synthesis is described. It is based on an articulatory model, which is controlled by seven parameters. The synthesizer employs a stylized vocal tract model that is realized using wave digital filter techniques. It includes a self‐oscillating glottis model, voiceless excitation in the tract, damping, a nasal tract, and a radiation load that allows the simulation of lip protrusion. The synthesizer runs in real time on a signal processor. The described model was fit to natural speech using a Kalman filter algorithm. Simple rules for the synthesis of unrestricted German speech are described.

Separation of several speakers recorded by two microphones (cocktail-party processing)

Abstract A signal processing method for enhancing the directional separation of an ordinary (dummy-head) stereophonic speech recording is described that, after initial adaptation to a certain direction, simulates the human ability to concentrate on speech coming from this direction and to suppress disturbing speakers from other directions. The method is derived from the principle of Adaptive Noise Cancelling; an FFT/Overlap-Add realization of the adaptive filter is chosen and short-time power estimates are used for its determination. In tests with up to four speakers, clear improvements of SNR and intelligibility of the desired speaker were obtained.

Recognition of isolated words based on psychoacoustics and neurobiology

Abstract A simple neural network for isolated word recognition constructed under consideration of neurobiological and psychoacoustical observations is described. The biologically motivated preprocessing of the speech signals includes transforming frequency to critical band-rate and power to loudness, contrasting the spectrograms and extracting temporal and spectral features. It is shown that the different stages of preprocessing of the speech signal increase recognition rates significantly and are essential to achieve faultless recognition of a small vocabulary. In addition, the network is able to recognize simultaneously spoken words without any change of its architecture. Thus, it represents a concept to solve one of the most difficult figure-ground-problems in speech research without using conventional techniques like directional separation of stereophonically recorded speech or fundamental frequency tracking.

Time-varying wave digital filters for modeling analog systems

The theory of wave digital filters, which are well suited to model analog networks including those containing transmission lines, is extended to the time-varying case. This allows simulation of the moving vocal tract or parametric amplifiers. A representation consistent with the physics of time-varying reactances and transmission lines requires not only time-varying filter coefficients, but also structural modifications in the n-port adaptors and reactance one-ports of the filter. The number of coefficients in the adaptors is increased, and stability and pseudopassivity are not assured.

Untersuchungen zur spezifischen Dauer deutscher Vokale

This paper examines the effects of vowel quality (formant frequencies), speaker and place of articulation of the surrounding consonants (/p/, /t/ or /k/) upon the duration of German vowels embedded in

Speech Recognition by Hierarchical Segment Classification

A neural network for speech processing is presented. Its complex architecture, incorporating selforganizing feature maps [1], allows the construction of a hierarchy of layers, where each layer operates on a larger time scale and deals with higher units of speech, like phonemes, syllables, word parts and so on. Tasks the network has to deal with include representation of speech, segmentation of the speech signal and classifying segments.

Linear prediction on a warped frequency scale (speech processing)

An ordinary predictor and a frequency warped predictor are compared in an ADPCM (adaptive differential pulse code modulation) system. The experimental results show that for an unwarped predictor of order ten, the order of the warped predictor can be reduced by two for the same speech quality. The audible differences between the normal and the warped predictors are very small, so the statements of the test subjects are widely spread. The measured noise of the system with the warped predictor is about 2 dB higher than the normal predictor, but it has lower audible noise. However, the plosives (p, t, etc.) are more blurred. The measured SNR was taken from the difference of the signals before the preemphasis filter and behind the deemphasis filter at the end of the system. Comparing the computing time needed to obtain the predictor coefficients, the difference between the normal and the warped predictor is negligible. >