Formant Frequency Estimation Research Articles

An application of mapping neural networks and a digital signal processor for cochlear neuroprostheses.

Cochlear neuroprostheses strive to restore the sensation of hearing to patients with a profound sensorineural deafness. They exhibit a stimulation of the surviving auditory nerve neurons by electrical currents delivered through electrodes placed on or within the cochlea. The present article describes a new method for an efficient derivation of the required information from the incoming speech signal necessary for the implant stimulation. Also some realization aspects of the new approach are addressed. In the new strategy, a multilayer neural network is employed in the formant frequency estimation having some suitable speech signal descriptors as particular input signals. The proposed method allows us a fast formant frequency estimation necessary for the implant stimulation. With the developed strategy, the prosthesis can be adjusted to the environment which the patient is supposed to live in. Moreover, the neural network concept offers us an alternative for dealing with the areas of neural loss or "holes" in the frequency map of the patient's ear.

A feedback control system for real-time formant estimation. I--Static and dynamic analysis for sinusoidal input signals.

This paper presents a novel analog scheme suitable for the real-time estimation of formant frequencies. Formant tracking is based on a feedback technique which uses both the amplitude and phase characteristics of two stagger-tuned bandpass filters to give an improved dynamic behavior. The implementation of the system requires a small number of components, and is practical for low-power applications. An analysis of the static and dynamic behavior is given for sinusoidal input signals. The transient response is independent of the amplitude level of the input signal. The system is designed for second formant detection in a cochlear prosthesis system.

Formant frequency estimates for abruptly changing area functions: a comparison between calculations and measurements.

Vocal tract area functions may contain quite abrupt changes in cross-sectional area. In formant frequency calculations for such area functions, an inner length correction (ILC) should be applied. The relevance of this correction was investigated by comparing acoustic measurements obtained from a physical model of the vocal tract with data gathered by means of computer simulations. Calculating formant frequencies without applying internal length corrections caused substantial errors, particularly for area functions representing apical stops just anterior to occlusion. Decentering and axial symmetry in the arrangement of the area elements of the physical model were briefly studied and found to have effects on the formant frequency values.

Zero-crossing based spectral analysis and SVD spectral analysis for formant frequency estimation in noise

The authors discuss a method for spectral analysis of noise corrupted signals using statistical properties of the zero-crossing intervals. It is shown that an initial stage of filter-bank analysis is effective for achieving noise robustness. The technique is compared with currently popular spectral analysis techniques based on singular value decomposition and is found to provide generally better resolution and lower variance at low signal to noise ratios (SNRs). These techniques, along with three established methods and three variations of these method, are further evaluated for their effectiveness for formant frequency estimation of noise corrupted speech. The theoretical results predict and experimental results confirm that the zero-crossing method performs well for estimating low frequencies and hence for first formant frequency estimation in speech at high noise levels ( approximately 0 dB SNR). Otherwise, J.A. Cadzow's high performance method (1983) is found to be a close alternative for reliable spectral estimation. As expected the overall performance of all techniques is found to degrade for speech data. The standard autocorrelation-LPC method is found best for clean speech and all methods deteriorate roughly equally in noise.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Homomorphic linear predictive coding: a new estimation algorithm for all-pole speech modelling

The authors present a new estimation algorithm for an all-pole model, named as homomorphic linear predictive coding (HLPC) which models the vocal tract transfer function as an all-pole filter, but performs the estimation of the filter coefficients in the cepstrum domain by a minimum mean squared error method. By limiting the summation interval of squared errors in the cepstrum domain to the low time portion that is not affected by pitch components, the estimation results obtained by HLPC are unbiased and independent of the exciting signal type. Experiments on spectrum estimation and formant estimation under several conditions have been carried out for comparison. It is shown that for pitch-asynchronous analysis, HLPC has a higher accuracy than LPC in spectrum estimation as well as formant frequency and bandwidth estimation, especially for speech signals with high pitch frequencies.

Formant frequency estimation by moment calculation of the speech spectrum

Moment calculation is applied to extract the formant frequencies of a speech spectrum. Three kinds of first‐order moments divide a spectrum into four frequency regions. The centers of gravity of the first three regions are calculated to give the 0th order estimation of the 1st, 2nd, and 3rd formant frequencies. Then the upper and the lower bounds of each region are modified so that the estimated frequency comes closer to the major peak of the spectrum, utilizing the second‐order and the third‐order moments that represent the variance and skewness of the spectral pattern. The process repeats until the k th estimation equals the (k − 1) th estimation. This modification improves the estimation precision significantly. An experiment with model spectra generated by an all‐pole model gave estimation precision of 3% using formant frequencies typical of the five Japanese vowels. Speech materials spoken by five male and five female speakers were used for this experiment. The speech waveform was sampled at a rate of 10 kHz through a 5 kHz LPF, quantized into 12 bits; then the spectrum envelope was calculated with the first 24 cepstra of a 256‐point FFT spectrum. The results give acceptable precision, compared with visually determined formant frequencies.

Analysis of speech signals of short pitch period by a sample-selective linear prediction

The conventional linear prediction analysis has difficulties in estimating the vocal tract characteristics of voiced sounds uttered by females or children. This paper shows that the vocal tract characteristics of those speech signals can be estimated accurately by the sample-selective linear prediction (SSLP) method proposed by the authors. The SSLP is a two-stage linear prediction analysis employing only relevant sample values in the second stage analysis, while the conventional linear prediction method employs all the sample values with equal weights as predicted values. The accuracy of the proposed method in estimating formant frequencies is examined on synthetic vowels of short pitch periods. The validity of the method is confirmed by inspecting the estimated spectral envelopes and distributions of the estimated formant frequencies of natural vowels uttered by a female.

Instantaneous Frequency Estimation Using Adaptive Linear Predictor Weights

A computationally efficient scheme for estimating the digital instantaneous frequencies of narrowband inputs is introduced. The frequency estimates are obtained by searching for minima of the inverse input power spectrum. This spectrum is estimated at each input sample from the weights of an adaptive linear predictor which uses the LMS (least mean square) algorithm to update its weights. The related minima are sought via an iterative search algorithm, referred to as the iterative frequency estimator. This algorithm is computationally more efficient than available methods, and also provides a higher resolution. Simulation results are included; these include tracking of random message sequences in FM signals, and the formant frequency estimation of speech.

Vowel normalization by frequency warped spectral matching

Normalization of formant frequencies have frequently been used to eliminate inter-speaker differences in vowel recognition. However, estimation of formant frequencies becomes difficult under certain circumstances, such as for telephone speech. This paper presents an approach to vowel normalization based on frequency warped spectral matching. A frequency normalized distance between test and reference spectra is defined on the basis of the minimum mean square difference over all possible choices of frequency warping functions under certain nonlinearity constraints and boundary conditions. After adaptively eliminating spectral slope differences due to the individual glottal characteristics, the spectral distance is computed by means of dynamic programming. The vowel identification experiments were conducted on the nine American English vowels in /hvd/ utterances spoken by 12 male and 12 female speakers. The results indicated that the frequency warping method substantially increased the identification scores for female vowels when the male vowels were used as reference. They also indicated that although the improvement in identification was attributed mainly to the linear frequency scaling, an additional improvement for vowel /ae/ was obtained by a slight nonlinear frequency warping. In addition, an application to speaker normalization for word detection in connected speech is discussed.

Formant estimation of high fundamental frequency speech

Formant measurement procedures often rely on there being a low fundamental frequency. An early study [B. Lindblom, International Congress of Phonetic Sciences, 4th, Helsingfors, 1961, 189–202 (1962)] found that the mean error in formant estimation ranged from about 40 Hz to a frequency of one‐fourth the fundamental. This study compares signal processing techniques for the estimation of formant frequencies and bandwidths of synthesized and natural speech characterized by a high fundamental frequency. Utterances were synthesized [D. H. Klatt, J. Acoust. Soc. Am. 67, 971–995 (1980)] using young children's utterances as models. The spectral and durational characteristics were matched closely by manipulating the synthesizer parameters. Spectrograms, discrete Fourier transforms, linear prediction envelopes, and auditory pseudospectrograms were computed for both the synthesized and natural utterances. The accuracy of formant estimation was judged by comparing the values determined by each of these methods to the known frequencies and bandwidths of the synthesized speech. Implications for formant estimation of natural speech will be discussed. [Work supported in part by a Whitaker Health Sciences Fellowship.]

A pretty good formant tracker

Algorithms were developed for estimating formant frequencies on a time frame by time frame basis. The algorithms incorporate acoustic phonetic knowledge about formants. The algorithms are based on knowledge of acoustic phonetics, and do not merely rely on spectral peaks. Thus two formants will be identified even when the first and second formants merge or when the second and third formants merge to form a single peak. The algorithms were evaluated by comparing formants drawn on a speech spectrogram by a trained phonetician to those drawn automatically. The data base consisted of both male and female speakers speaking phonetically balanced sentences. In addition, synthesis derived from the formant frequencies and amplitudes was tested for intelligibility. Results will be presented for these two procedures. [Supported by NSF and DARPA.]

Accuracy of formant frequency estimation by spectrograms and by linear prediction analysis

The accuracy of spectrographic techniques and of linear prediction analysis in measuring formant frequencies was compared. The first three-formant frequencies of ninety synthetic speech tokens were measured by three experienced spectrographic readers and by linear prediction analysis. The synthetic speech tokens were made by parallel synthesis and were chosen to represent a wide range of formant bandwidths, frequencies, and fundamental frequencies. For fundamental frequencies between 100 and 350 Hz, both methods are accurate to within approximately ±60 Hz for both first and second formants; the third formant can be measured with the same degree of accuracy by linear prediction, but only to within ±120 Hz by spectrographic means. For linear prediction analysis, accuracy does not decrease with increasing fundamental frequency, although the accuracy of both methods decreases greatly when fundamental frequency is 400 Hz or greater. These limits of measurement appear to be within the vicinity of difference limens for formant frequencies. [Supported by NINCDS Grant NS03856.]

Influence of the position of the analysis frame in LPC pitch-synchronous analysis

The paper deals with the effect of the position of the time interval on the evaluation of voiced sound characteristics in speech analysis using the Linear Prediction Coding (LPC) technique. It is shown that when the analysis frame coincides with a single pitch period (pitch-synchronous analysis) an erroneous alignment (with respect to pitch pulses) of the analysis interval may introduce significant errors in the estimation of formant frequencies and bandwidths, and more generally, of sound spectrum. Synthetic speech is used to investigate the phenomenon and the experimental results are discussed. On the basis of these results, various techniques for reducing the influence of the position of the analysis interval on Linear Prediction parameters are discussed.

Spectral smoothing technique in PARCOR speech analysis-synthesis

In linear predictive analysis of speech, voice periodicity influences formant frequency and bandwidth estimation accuracy. One of the most serious errors in estimating formant parameters is bandwidth underestimation that causes a quality difference between synthetic and natural speech. In this paper, the spectral smoothing technique (SST), using a lag window, is introduced in an autocorrelation method of the linear predictive analysis. In order to assess the effectiveness of the SST to reduce estimation errors, experimental comparisons of the usual autocorrelation method and the SST are presented. Spectral sensitivity analysis is also presented to evaluate the SST from the viewpoint of parameter quantization properties. SST features are summarized as follows: 1) Bandwidth underestimation elimination. 2) Spectral sensitivity reduction of PARCOR coefficients. 3) Simplicity in hardware implementation.

Vowel normalization by frequency warping

Normalizations of formant frequencies have frequently been used to eliminate inter-speaker differences in vowel identification. However, estimation of formant frequencies become difficult under certain circumstances. This paper proposes a new approach to vowel normalization based on a frequency warped spectral matching. A frequency normalized distance between test and reference spectra is defined on the basis of the minimum mean square difference over all possible choices of frequency warping functions. The warping functions are constrained so that they mostly take account of linear scaling for the differences in vocal-tract lengths and of slight nonlinear differences. The slope of the warping function is restricted by an assumed range of the vocal-tract length. After adaptively eliminating the individual glottal characteristics, the spectral distance is computed by means of dynamic programing. A preliminary experiment indicates a substantial improvement in identifying female steady-state vowels when male vowels are used as references. The vocal-tract length can also be estimated from the slope of the warping function.

Formant tracking algorithm for a speech analysis-synthesis system

A formant tracking algorithm has been developed to provide control signals automatically for a parallel formant speech synthesizer. In voiced sounds, the algorithm first estimates positions of high-energy excitation in the sampled speech waveform to be analyzed. These positions generally correspond to the points of glottal closure, and the interval following each position should correspond to a closed glottis region. For each such position the speech samples in the following few milliseconds are used to calculate a log power spectrum. An analysis-by-synthesis procedure iteratively updates estimates of formant frequencies and amplitudes using spectral matching criteria. A similar spectral-matching procedure is used during unvoiced sounds, but time averaging is introduced as appropriate when deriving the spectra to be matched. To cope with situations when the formant structure is not clearly defined in the power spectrum the iteration is performed separately for several different allocations of formants to spectral peaks. A final choice of formant frequencies and amplitudes is made every 10 msec and depends on both goodness of spectral match and continuity of formant tracks. Demonstrations will be made of the speech quality achieved for a number of male talkers.

Acoustic characteristics of speech produced with an electronic artificial larynx

This study investigated the acoustic characteristics of speech produced with an electronic artificial larynx (electrolarynx). Five normal-speaking adult males were taught to use an electrolarynx, and each talker recorded a set of speech materials in his normal voice and with an electrolarynx. The recordings were subjected to analyses which included: (1) intensity level measurements of speech as well as the direct radiated sound from the electrolarynx; (2) speech-duration measurements; (3) spectral analysis of the radiated masking sound from the electrolarynx; and (4) formant frequency estimates derived from the spectrum of a linear prediction inverse filter. Results indicated that the overall intensity of electrolarynx speech was similar for the five talkers, but that intensity of the radiated noise, and hence speech-to-noise ratios, varied considerably among talkers. In addition, talkers required about 50% more time to produce speech with an electrolarynx, indicating modification of articulatory behavior. Spectral composition of the radiated sound from the electrolarynx varied from talker-to-talker. Formant frequencies for a given talker were similar for the normal voice and the electrolarynx productions, but the intensity relationships between formants was different in the two modes of production. [Research supported by NINCDS and a grant from the Samuel W. Traylor Foundation.]