Formant Tracking Research Articles

A zero-crossing consistency method for formant tracking of voiced speech in high noise levels

A technique is described for the determination of the first three formant frequencies of a voiced speech signal corrupted by high levels of noise. The technique involves a rough frequency-domain analysis followed by an accurate time-domain statistical analysis. The frequency-domain analysis is performed with a bank of overlapped, shifted, bandpass digital filters chosen to cover the frequency range of the first three formant frequencies. The time-domain analysis involves a calculation of the consistency of the intervals between successive zero-crossings at the output of each filter. As the algorithm proceeds, those filters which contain peaks in the energy spectrum are chosen as a first estimate of the location of the first three formants. The final choice of filter containing each formant is made, in the neighborhood of the first estimate, by choosing that filter with a signal output having the maximum zero-crossing interval consistency. A decision procedure is applied to several calculated parameters to result in a measurement of each of the first three formant frequencies. The results of extensive testing of the method are reported for speech contaminated by high levels of additive white Gaussian noise.

An algorithm for formant tracking

A formant tracking algorithm than first forms strings of spectral peaks considering only the relative positions of the peaks and then assigns these strings to the formants according to the relative positions and lengths of the strings is presented. An example is also given.

Acoustic and articulatory asymmetries in fricative-vowel and vowel-fricative productions

When synthetic noise replaces the natural friction in FV syllables and VF syllables, the influence of the vocalic segment on the perception of the noise as /s/ or /∫/ is greater for the FV syllables [S. D. Soli and V. A. Mann, J. Acoust. Soc. Am. Suppl. 1 71, S75 (1982)]. The acoustic and articulatory bases of this perceptual asymmetry were examined. Mean formant tracks displaying confidence intervals and amplitudes were computed for natural FV and VF syllables, where F = /∫/ and /s/, and V = /ɑ/ and /u/. FV and VF transitions were equally stable, although FV transitions exhibited larger coarticulatory differences due to fricative context. FV transitions also occurred during the highest amplitude portion of the vocalic segment, while VF transitions occurred as voicing amplitude was decreasing rapidly before the upcoming voiceless fricative. These acoustic differences would lead to larger vowel context effects for FV syllables. Thus the perceptual asymmetries appear to be a consequence of differences in the organization and timing of the supralaryngeal and laryngeal events which produce initial and final voiceless fricatives. [Supported by NICHD.]

Vowel distortion in traumatic dysarthria: a formant study.

The frequencies of the first two formants of three German vowels in word context were determined by a formant tracking routine for 8 male subjects with closed head trauma. A centralized formant pattern was found to characterize vowel articulation in traumatic dysarthria. Different degrees of severity could be assessed and the process of recovery was described by the parameter 'formant triangle area'. An interpretation in articulatory terms is given and the results are discussed in the light of articulatory models.

Adaptive tracking of multiple sinusoids whose power levels are widely separated

The behavior of the gradient transversal filter (LMS), the gradient lattice (GL), and the least squares lattice (LSL) when used to track multiple sinusoidal (or narrow-band) components whose power levels are widely separated is investigated. These approaches to the realization of the pth-order one-step linear predictor of the time series are recursive in time. The lags of the instantaneous frequency estimates from their actual underlying values are of particular interest. The frequency tracking characteristics of the LMS, GL, and LSL algorithms are illustrated in several situations. Included are simulations of dual sinusoids undergoing a variety of frequency versus time dynamics and a formant tracking example of real speech.

Adaptive tracking of multiple sinusoids whose power levels are widely separated

The behavior of the gradient transversal filter (LMS), the gradient lattice (GL), and the least squares lattice (LSL) when used to track multiple sinusoidal (or narrow-band) components whose power levels are widely separated is investigated. These approaches to the realization of the pth-order one-step linear predictor of the time series are recursive in time. The lags of the instantaneous frequency estimates from their actual underlying values are of particular interest. The frequency tracking characteristics of the LMS, GL, and LSL algorithms are illustrated in several situations. Included are simulations of dual sinusoids undergoing a variety of frequency versus time dynamics and a formant tracking example of real speech.

A formant tracking system toward automatic recognition of speech

This paper describes the implementation of a Speech Understanding System component which tracks the formants of pseudo-syllabic nuclei containing voiced consonants. The nuclei are isolated from continuous speech after a precategorical classification in which feature extraction is carried out by modules organized in a hierarchy of levels. FFT and LPC spectra are the input to the formant tracking system. It works under the control of rules specifying the possible formant evolutions given previously hypothesized phonetic features and produces fuzzy graphs rather than usual formant patterns because formants are not always evident in the spectrogram pattern.

Speaker identification: New vectors for SAUSI

A semiautomatic speaker identification system (SAUSI) has been described (Proc. IEEE Conf. ASSP, 768–771 (1977)). This system employs four major vectors: (1) long-term speech spectra, (2) f0 analysis, (3) formant tracking, and (4) temporal analysis. Recently, the f0 analysis and temporal analysis vectors have been replaced by new ones each consisting of an expanded number of parameters. The new f0 vector is based on the mean and s.d. of the frequency distribution plus a parameter array consisting of the numbers of frequencies falling into 15–20 semitone interval “bins.” The temporal vector consists of an array of four 2–34 parameter subvectors as follows: (1) duration of speech rate (DSR), (2) voice/voiceless speech-time ratio (V/VL), (3) time-energy distribution (TED), and (4) vowel/consonant duration ratio (V/C). The nature and analysis techniques of these several vectors will be discussed as will their robustness as speaker identification cues in studies involving speaker and system distortions.

Specific contributions of the ARPA SUR project to speech sciences

This critical review of the recent ARPA Speech Understanding Research (SUR) project emphasizes the basic speech science results (only some of which were incorporated into the final systems). ARPA SUR provided significant progress in such parameter extraction methods as pitch detection and formant tracking. Advanced methods of speech segmentation and labeling ranged from allophonelike acoustic templates to lattices of alternative phonetic units, to quasiphonemic labeling, with emphasis on manner classes rather than place of articulation. The phonological tie between small- and large-unit recognition involved tradeoffs between extensive sets of alternative lexical baseforms and large sets of phonological rules. The syllable was promoted as a more basic unit than the phoneme, but difficulties arose with the use of syllable boundaries in baseforms and rules. Prosodic features were demonstrated as useful for determining sentence structure and confirming lexical hypotheses, but were not fully tested within any system. ATN grammars, semantic networks, and breadth-first parsers were effective for speech analysis, provided they were used to constrain system choices rather than simply be good models of language. [Work supported by Tri-Services and ARPA.]

Modifications to formant tracking algorithm of april 1974

This correspondence describes an improved version of the linear prediction formant tracking algorithm which was described in [1]. The new algorithm, like the original one, applies continuity constraints and branches out from an anchor point in the middle of each vowel. The changes are that the initial estimates for the formant frequencies at the anchor are determined more carefully, and that the option of choosing a new anchor point near the original one is allowed if the original one caused problems.

A pitch-synchronous digital feature extraction system for phonemic recognition of speech

The system described in this paper is subdivided into three main steps: pitch extraction, segmentation, and formant analysis. The pitch extractor uses an adaptive digital filter in time-domain transforming the speech signal into a signal similar to the glottal waveform. Using the levels of the speech signal and the differenced signal as parameters in time domain, the subsequent segmentation algorithm derives a signal parameter which describes the speed of articulatory movement. From this, the signal is divided into "stationary" and "'transitional" segments; one stationary segment is associated to one phoneme. For the formant tracking procedure, a subset of the pitch periods is selected by the segmentation algorithm and is transformed into frequency domain. The formant tracking algorithm uses a maximum detection strategy and continuity criteria for adjacent spectra. After this step, the total parameter set is offered to an adaptive universal pattern classifier which is trained by selected material before working. For stationary phonemes, the recognition rate is about 85 percent when training material and test material are uttered by the same speaker. The recognition rate is increased to about 90 percent when segmentation results are used.

Speaker-identifying features based on formant tracks.

The formant structure of three diphthongs, four tense vowels, and three retroflex sounds was examined in detail for possible speaker-identifying features. These sounds were spoken five times each in sentence context by ten speakers of General American on one day and by six of the speakers on a second day at least three weeks later. Formant tracks were computed for each sound under investigation using covariance-type pitch-asynchronous linear prediction together with a root-finding algorithm. The interspeaker variability of about 200 measurements made on these formant tracks was compared initially with intraspeaker variability through the calculation of F ratios. Those with average F ratios greater than 60 were evaluated further with a probability-of-error criterion. Features that are potentially most effective in identifying speakers are the minimum second-formant value in [-r], the maximum first-formant value in [-r], the maximum second-formant values of [o] and [-I], and the minimum third-formant value of [-]. The individual differences apparent in these sounds presumably depend more on speaker habits than on vocal-tract anatomy. The error bound predicted for a speaker identification procedure based on these five features is 0.24%. An identification experiment using only the best two features gave 12 errors out of 80 identifications. Subject Classification: [43]70.65, [43]70.40.

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.

Automatic tracking of speech parameters for parallel formant synthesis

A speech synthesis model in which the formant resonators are connected in series does not produce satisfactory speech quality for sounds produced with a narrow constriction in the vocal tract or with the velum lowered so that the nasal cavity is coupled to the oral cavity. These latter cases are better handled by a parallel connection of formants, in which the amplitudes of the individual formants can be controlled independently. An existing analysis-synthesis method for a series model of speech [J.P. Olive, “Automatic Formant Tracking by a Newton-Raphson Technique,” J. Acoust. Soc. Am. 50, 661–670 (1971)] has been adapted to a parallel formant model of speech. The analysis procedure derives pitch and four formant frequencies and amplitudes from natural speech which are then used to control a speech synthesizer. The quality of synthetic speech produced in this manner is a substantial improvement over that produced according to a series model.

A system for acoustic-phonetic analysis of continuous speech

A system for acoustic-phonetic analysis of continuous speech is being developed to serve as part of an automatic speech understanding system. The acoustic system accepts the speech wave-form as an input and produces as output a string of phoneme-like units referred to as acoustic phonetic elements (APEL's). This paper should be considered as a progress report, since the system is still under active development. The initial phase of the acoustic analysis consists of signal processing and parameter extraction, and includes spectrum analysis via linear prediction, computation of a number of parameters of the spectrum, and fundamental frequency extraction. This is followed by a preliminary segmentation of the speech into a few broad acoustic categories and formant tracking during vowel-like segments. The next phase consists of more detailed segmentation and classification intended to meet the needs of subsequent linguistic analysis. The preliminary segmentation and segment classification yield the following categories: vowel-like sound; volume dip within vowel-like sound; fricative-like sound; stop consonants, including silence or voice bar, and associated burst. These categories are produced by a decision tree based upon energy measurements in selected frequency bands, derivatives and ratios of these measurements, a voicing detector, and a few editing rules. The more detailed classification algorithms include: 1) detection and identification of some diphthongs, semivowels, and nasals, through analysis of formant motions, positions, and amplitudes; 2) a vowel identifier, which determines three ranked choices for each vowel based on a comparison of the formant positions in the detected vowel segment to stored formant positions in a speaker-normalized vowel table; 3) a fricative identifier, which employs measurement of relative spectral energies in several bands to group the fricative segments into phoneme-like categories; 4) stop consonant classification based on the properties of the plosive burst. The above algorithms have been tested on a substantial corpus of continuous speech data. Performance results, as well as detailed descriptions of the algorithms are given.

A control program for a Glace-Holmes speech synthesizer

Synthetic speech provides unique opportunities to investigate important acoustic cues in speech signals and the natur~ of speech perception in general. The program, SYNG, descnbed here is a utility program which controls a parallel-resonance terminal-analog speech synthesizer JAWORD (Glace, 1968). The synthesizer requires up to 11 inputs, 1 of which determines types of excitation (noise source, pulse source, or combination of the two), 1, fundamental frequency, and the remaining 9, the formant frequencies and their amplitudes. Synthesizer control by computer essentially involves (1) preparation of sets of 11 parameter values, each set representing one time segment. of speech, and (2) sending these sets of values to the synthesizer through digital-to-analog converters under programmed timing control. Input/Output. Input data, from a card reader, Teletype,. or magnetic tape unit, specify each set of 11 values corresponding to a 10-msec sample of speech. An array of parameters for 600 time samples, which is equivalent to 6 sec of speech, can be accommodated at one time. With the use of disk storage, 20 such parameter arrays can be stored and synthesized in any order, extending the system capacity to handl~ up to 120 sec of ~peech. An option is also available to synthesize SOUGHS (Society of Users of Glace-Holmes Synthesizer) data punched on cards. When parameter values of some time segments are not available, the program provides either linear or exponential interpolation to generate the missing information in a manner similar to that described by Rabiner (1968). The program then converts the successive sets of parameter values into equivalent voltages that are transferred to the synthesizer. The transfer rate can be specified at the time of execution, so that the synthesized speech can be compressed or stretched in time. A silent interval between the repetitions is also specifiable. On-going synthesis halts upon detection of an interrupt given manually by pressing a control button at which time the user is given the several procedural options, For example, he can listen to any portion of the entire utterance at any desired transfer rate, he can obtain a printout of arrays of parameter values for any portion, or he can edit one or more parameter values for any segment. The edited portion can be synthesized alone or in the context of the entire utterance, or its parameter arrays can be written out and/or remodified. If desired, parameter values can be displayed as a function of time on an oscilloscope (one parameter) or on a graphic level recorder (up to eight parameters). The program is written in an interactive form so that instructions to the user are typed out on the Teletype preceding each action required, and messages are given when illegal actions are taken. In contrast to synthesis by rule, in which input data typically specify a sequence of phonetic elements and accent and intonation codes (Mattingly, 1968), the present program utilizes parametric descriptions of utterances supplied by a primary recognition program (Hughes et al, 1969) which includes formant tracking and fundamental frequency tracking algorithms (Snow &.Hughes, 1969). Program Language and Computer. The program is written in FORTRAN IV for a CDC 1700 computer with 16 channels of digital-to-analog conversion capability. Availability. The listing of the program is available free on request from the author. REFERENCES

Auditory-motor formant tracking: a study of speech imitation.

Auditory-motor formant tracking, or the vocal reproduction of formant patterns, is one aspect of speech imitation skill. The study reported here assessed the ability of four adult speakers to imitate synthesized vocalic stimuli. These stimuli took the form of two steady-state segments joined by a transitional segment. The first steady-state segment corresponded to one of eight American English vowels, and the second, to one of 14 vowels that were not expected to have a prominent phonemic identity in the language. Spectrographic analyses of the imitative responses allowed comparisons of the formant structure for the synthesized stimuli and the corresponding human reproductions. Analyses of the spectrograms revealed that the directions of movement for the first two formants were almost always reproduced accurately, but the extent of movement frequently was overshot. These responses were judged to be consistent with a contrast effect in speech perception, a phenomenon previously discovered in experiments on vowel identification. The variability of formant reproduction for a given vowel was predicted at least roughly by the ambiguity of the stimulus in a preliminary identification experiment. These results suggest that the responses in an imitation task are intermediate in dimensionality to the responses in discrimination and identification tasks.

Formant synthesis by rule in a speech understanding system

A speech understanding system is a device that will accept and demonstrate an understanding of spoken English sentences within limited syntactic, semantic, and lexical domains. Naturally spoken sentences involve the recoding of phonetic segments by a process that is most easily described by a set of generative phonological rules, many of which are optional. Thus phonological processes such as assimilation, reduction, flapping, and deletion, etc., can be accounted for optimally by designing a speech understanding system component that does varification of an hypothesized word or word string. The effects of segmental coarticulation on formant motions and on other acoustic parameters can also be treated conveniently in a generative mode by incorporating a set of rules for formant synthesis into the word-varification component. A set of rules will be described that generates formant tracks from an abstract phonological description of an utterance. The output of the rules has been compared with formant data extracted from a 1000-segment paragraph read by a single speaker. The adequacy of this schema for multiple-speaker word varification is discussed.

Basic directions in automatic speech recognition

This paper represents a view of basic problems in automatic speech recognition drawn from the vantage point of applied linguistics. The basic areas of speech production, articulatory phonetics, acoustic analysis, acoustic phonetics, and phonetic sequences of natural speech are crucial to flexible automatic speech recognition, especially for the long-range goal of recognizing continuous-flow large-vocabulary natural speech of different speakers. Articulatory phonetics provides a link between the physical events of speech and the elements of the phonological code. The processes of speech production are basic to articulatory phonetics while the acoustic theory of speech production provides a basis for the phonetic interpretation of the acoustic. speech waveform. Formant frequencies are particularly important acoustic parameters and recent improvements in formant tracking are encouraging for speech recognition. Even if complete phonetic recognition is realized, however, there remains the conversion of phonetic strings to lexical strings. This problem requires extensive phonetic analysis of actual spoken language.

Automatic formant tracking by a Newton-Raphson technique.

A wide range of speech investigation requires natural-sounding synthetic speech whose individual acoustic features—such as pitch, formants, and duration—can be varied independently. This study describes an algorithm for obtaining such synthetic speech by automatically analyzing natural speech to obtain data for the control of a formant synthesizer. The lowest three formants of natural speech were determined by finding simultaneous solutions of the least-square-fit equations by means of a Newton-Raphson technique. The data obtained from this formant analysis, together with pitch data which were also obtained automatically, were used for controlling a computer-simulated formant synthesizer. This algorithm was tested for various sentences and speakers, and the synthesized speech was found to be of good quality.