Measuring voice in the clinic-Laryngograph Speech Studio Analyses

Adrian Fourcin¹ Julian McGlashan² Richard Blowes³ 

¹University College London, England; ²Queens Medical Centre, 

Nottingham, England; ³Laryngograph Ltd., London, England 

e-mails  a.fourcin@ucl.ac.uk; julian.mcglashan@nottingham.ac.uk;lx@laryngograph.com 

Summary

The aim is to get down to simple basics  and to provide a straightforward set of clinically useable quantitative acoustic analyses that reflect important aspects of what is so obvious to the ear of the listener.

Loudness, pitch and quality are widely used to describe essential aspects of a speaker’s voice. However, their basic links with simple parameters of auditory processing are currently little used either in analysis or therapy. In the present work the application of pitch perceptual criteria is described in the provision of an integrated framework for practical clinical assessment and therapy. Sustained sounds are shown to require quite different levels of sampling accuracy from those needed for ordinary connected speech. Radically useful results come from the application of these basic ideas to the representation and analysis of the pathological and the normal voice in case management and audit. Even very abnormal speech samples can be seen to have measurable structures of pitch, loudness and quality  in the midst of apparent disorder.

Examples are discussed of the application of these approaches to pathological voice samples from clinics in several countries. Results from five main analysis types are examined:

    •  sustained vowel measurement using standard techniques but with 1MHz period time sampling 

    •  vocal fold frequency distributions based on voice frequency difference limen related bin sizes and both first and second order (digram) analyses to show the effect of pitch perturbation 

    •  crossplots of vocal fold period to period variability which give an overview of intrinsic structure and provide a base for the measurement of irregularity which takes account of normal intonational variation in the speaker’s voice 

    •  phonetogram and amplitude distribution analyses of connected speech using both simple and digram plots for ordinary connected speech 

    •   closed phase ratio distributions again using first and second order analyses to provide measures of this aspect of voice quality.

Two signal inputs have been used, from an acoustic microphone and an electrolaryngograph. These inputs are also basic to a related development in stroboscopy. Although sustained sounds depend on different mechanisms of auditory monitoring and productive control, data from connected speech analysis can be of vital help in guiding the design and clinical use of new stroboscopic equipment.

Using pitch perception to guide voice measurement

For most practical purposes the really important aspects of voice are those that can be heard, and the dominant dimension in hearing voice is pitch. This simple concept leads to the possibility of using some simple quantitative criteria to detect and quantify the differences between “good” and “bad” voices.

Classically, pure tones provide a basic reference for both the definition and perceptual investigation of pitch. Subjective psychophysical data have been stably established over many years. Maximum discriminability is reached between 1 kHz, C6, near the top of the soprano register, and 2kHz with an average best just noticeable difference, jnd, of about 0.7% at 200Hz and 0.4% or 4 Hz in the region of 1kHz [1,2] with individual jnd sensitivities going down to 0.1%. Auditory pitch detection for the frequency ranges of the speaking and singing voice appear to employ mechanisms which operate on the basis of temporal processing [4]. This level of pitch discrimination implies an average ability to detect temporal differences between successive periods of about 4 µs, and for some individuals, 1µs. This temporal signal processing ability for pitch perception is paralleled in auditory lateralisation where interaural time differences of about 2µs to 10µs are detectable.

For steady complex tones and vowels in the fundamental frequency range of conversational speech, the pitch discrimination jnds are even smaller than those obtained with pure tones. Wier [3] and Moore [4] for example, within the range 200 to 600 Hz, reported jnd values from about 0.15% to 0.3%. For vowel-like sounds with simple changing fundamental frequency contours, however, the ability to perceive differences in fundamental frequency is drastically reduced and the jnd may be 8% at about 100 Hz [5]. This increase in, and magnitude of, jnd has also been found for whole word utterances with simple intonation contours, the jnd here never being less than 6%. When more complex contours are used, the differences needed to achieve reliable detection may be as great as 20% [6]. The subjective results for these sound types are not as well established as for sustained sounds and there is a dependence on the duration of the tone. There is, however, a good working consensus between a large number of reported observations [t’Hart et al]. These established observations give clear implications in respect of the accuracy criteria which should be aimed at for the analysis of the separate categories of sustained sounds and connected speech.

A basic set of tools for accurate voice pitch measurement 

Tool number 1 

Most methods of voice pitch analysis depend on the use of the acoustic signal of speech sampled at low rates which do not correspond to the requirements imposed by the pitch dL performance of the ear. The essential need which has to be met is illustrated by the graph below.

Figure 1 Voice pitch measurement error and sampling rate 

At a voice frequency of 1000Hz in the soprano range the human ear can detect changes of around 0.1%. In order to do as well as this it is necessary to use a sampling frequency of 1MHz. This is what is done for the following measurements.

Tool number 2 

A second basic problem associated with conventional approaches to voice analysis comes from the inadequacy of pitch extraction algorithms based on the acoustic signal. A more reliable technique is to use the electrolaryngographic [egg] output from the speaker’s voice activity.

Figure 2 electrolaryngograph voice signal, Lx, [egg] sample with period markers

The figure shows a standard Lx/egg waveform with period markers superimposed. These markers give the basis for an accurate determination of each individual pitch period, Tx, and they can be sampled at 1MHz to give a basis for measurement which, although very highly detailed for many purposes, is linked to the best that the ear can do and provides for considerable flexibility in computational manipulation.

It is, of course, quite easy to deal with sustained sounds but the method must also be reasonably robust when applied to the rapidly changing waveform of running speech. A typical output for a practical system is shown below.

Figure 3 Voice period, Tx, markers for changing pitch and larynx height

The figure shows the same process of marker generation as above, in Figure 2, but now the speech acoustic waveform, Sp, is included and the excerpt is from a sample of fluent speech. At the top of the figure the period by period “instantaneous” frequency, Fx, is shown, with an update for every vocal fold period. The vertical thickness corresponds to loudness being determined by the peak acoustic pressure for each period  a technique which is very useful for interactive clinical work in voice training.

Figure 4 “Pitch and Loudness Patterns” for a normal, left, and an abnormal voice, right.

The use of period by period sampling gives a very clear view of the difficulties that may be encountered by a speaker with a voice disability and shows the remarkable precision of normal voice pitch control. This type of precision data analysis is also basic to the provision of measurements both for sustained sounds and for the analysis of running speech.

Figure 5 Patterns for sustained vowel sounds produced by normal, A, and pathological, B, voices

The current clinical techniques for the quantification of voice abnormality depend to an appreciable extent on the use of sustained sounds and the standard protocol uses the steady state in the centre of the sample. Period by period analysis gives a clear indication of the onset and offset transients which this approach misses. The pathological speaker in general has difficulty in producing smooth voice onset and offset. This is clearly seen initially, where diplophonic breaks in the voice precede more steady production, and in the voice breaks at the end.

Table 1 productive jitter and perceptual dL comparisons

As must be expected, perception leads production but it is striking and commonly observed, that the pathological voice does not have a jitter commensurate with the disability. This relatively small difference results from the choice of the centre interval of a sound sustained at a comfortable pitch  and the speaker’s auditory monitoring ability for sustained sounds.

Connected speech and sustained vowels 

For the majority of the population, speech communication is at the heart of our daily lives. Clinical voice measurement, however, is mostly directed towards the appraisal of the ability to produce a sustained vowel. Since there are quite substantial perceptual differences between our ability to hear pitch regularity in sustained vowel sounds as opposed to fluent speech, it would be of interest to make at least an initial appraisal of the ways in which perception and production may interact in the voice pitch structures of the two types of phonatory activity. There may additionally be an advantage in comparing pitch regularity inspired analyses based on the two types of spoken material simply with a view to contributing to filling the gap between clinical indices of severity of dysphonia based on vowel measurement and those using a perceptual evaluation of continuous speech. Most important of all, however, is both to make use of pitch criteria and to take account of the nature of pitched sounds. Regular repetition of an acoustic event and perceived pitch go hand in hand.

Analyses of ordinary running speech

Figure 6 Vocal fold frequency,Fx, distribution for a 2m sample of normal connected speech

This figure shows a standard result for the analysis of the range of vocal fold frequencies contained in an ordinary sample of running speech. The analysis is not, however, standard because it is based on the period by period measurement of vocal fold frequencies with no smoothing. This leads to the possibility of showing and measuring the extent to which the voice contains well defined pitch components, because individual periods can be compared with each other.

Figure 7 Fx distributions for a whole normal sample (red) AND for its “pitched” components

The use of accurate period by period information makes it easy to plot the occasions when two successive pitch periods have essentially the same value. For the normal voice this happens very often indeed. The pathological voice, however, is very easily identified by the ear as having period to period irregularity. This important feature is shown in the inner of the two distributions above. The two distributions together give an immediate insight into important aspects of voice quality  pitch height and range, modal structure, and regularity. The technique is especially useful for pathological voice analysis.

Figure 8 Fx distributions for a whole abnormal sample (red) AND for its “pitched” components

The two distributions here are very dissimilar. DFx1, the outer plot, shows the distribution of Fx values for every vocal fold period in the whole 2m. sample. DFx2 shows only those Fx values for which two successive periods have been essentially the same. Two modes of vocal fold vibration are shown. The main at about 200 Hz is well defined. At about an octave below, the lower mode is more diffuse and is evidently associated with considerable period to period irregularity since the values of DFx1 and DFx2 are so different. Different pathologies give rise to different types of modal structural differences but for most cases the presence of voice pathology will be associated with marked discrepancies in magnitude and shape between these two forms of representation.

Jitter and irregularity in connected speech

Figure 9 Vocal fold period crossplots, CFx speaker A on the left, B on the right

Jitter and intonation

The procedure basic to the ordinary application of the jitter criterion is applied only to sustained sounds and requires that the voiced sound being measured is held at as constant a pitch as possible by the speaker. The essential concept, however, is directed at obtaining a quantitative assessment of pitch variability. The idea is just as applicable to ordinary connected speech so as to get an appraisal of the irregularity which may be inherent to the social use of a pathological voice.

An obvious first approach to the measurement of pitch irregularity in a sample of running speech is to determine the standard deviation of the spread of cycle to cycle differences in regard to periods or frequencies. A difficulty with this approach is that it will necessarily include ordinary intonational variations as part of the estimate of irregularity. The problem is perhaps best illustrated with reference to actual data. When vocal fold vibration is essentially regularly periodic the use of a period by period crossplot, as in Figure 9 A, gives a clearly defined diagonal line  since successive periods have almost the same values, apart from the variations arising from the intonational frequency related changes of connected speech. For the pathological voice, however, the shape of the crossplot is not so simply defined because successive vocal fold periods are very often markedly different and are not totally under the speaker’s cognitive control. This method of plotting the range of variability in period to period coherence is effectively similar to the application of the jitter criterion, used for sustained sounds, to the whole of a connected speech sample.

The interpretation of jitter in running speech, however, is not at all the same as that for sustained sounds. First, the pitch dLs are quite different in the two cases. The bin sizes needed for the adequate representation of significant changes in the present data involves 6% steps. The 0.1% resolution required for the analysis of sustained sounds is not appropriate. Second, the presence of intonational changes makes it necessary to ignore variations which are part of the normal patterning of vocal fold frequency change in running speech. Figure 9 A shows that there is indeed a centre continuous core of variation for the whole of the vocal fold frequency range and this is found for all normal speakers.

Figure 10 How the core of the normal Crossplot, CFx, is shaped by normal intonation

If the pitch difference limen value of 6% is applied to this data then it becomes possible to apply a theoretically founded criterion which makes it practically feasible to separate the variability arising from intonation from that due to other causes. It is then only necessary to determine all the pitch deviations which are more than 6% away from the centre line in the graph showing Fx1 against Fx2  where Fx1 is the frequency value of the first vocal fold cycle in any pair of cycles in the whole utterance and Fx2 is the frequency value of the immediately following cycle of the pair. Fx is used to denote the frequency value of a single vocal fold cycle, the period of this cycle being measured from point of closure to point of closure.

Normal and pathological voice examples

The comparison of Figure 9A with 9B shows how the relatively small jitter differences for the sustained sounds produced by these speakers in Table 1 is related to quite marked structural changes in their samples of connected speech. In these particular instances, irregularity is 3·2% for the normal speaker and 14·7% for pathological speaker B. Both values were measured in the way described above as a percentage of the number of vocal fold periods, outside the centre core of intonation-dependent pitch change, relative to the total number of vocal fold periods in the whole spoken sample.

Loudness and Quality 

Connected speech phonetogram

The standard phonetogram was designed to provide an overview of the dynamic range of a singer’s voice and was based on the separate production of sustained sounds. The same principle can be applied to the analysis of the speaking voice to give a “Dynamic Phonetogram” which is derived from the amplitudefrequency analysis of a complete sample of connected speech.

Figure 11 Dynamic Phonetograms derived from samples of connected speech: 

normal speaker A on the left; abnormal voice speaker on the right

In both Figure 11A and B, the first and second order distributions have been superimposed. This has little effect on the presentation for speaker A; it does have a profound influence on the form of the data presentation for speaker B since the presence of a bimodal peak in loudness is very evident in the first order distribution but not in the “pitch” related second order plot.

Figure 12 Period by period amplitude crossplots  CAx

A factor contributing to our perception of hoarseness comes from the irregularity of successive amplitude peaks in the cycle to cycle excitation of the vocal tract. This is especially evident in connected speech and speaker A on the left has a smaller spread in these analyses than B. Using a similar measure of irregularity to that employed for CFx gives values respectively of 3·3% and 6·5%.

Voice quality, “closed” phase and pitch

Figure 13 Using Lx transconductance to estimate “closed phase” ratio

A direct method for the estimation of the closed phase of each glottal cycle can be based on the use of the Lx/egg waveform. The black bars in Figure 13 show the closed phase at a point 70% down from the peak of the waveform and this is simply taken as a ratio with the total period, Tx, to give this parameter. In what follows the symbol Qx is used to denote this particular closed phase estimate. As before, Qx can be measured for each individual period and it can also be linked to the pitch regularity which occurs when two successive periods have essentially the same value  to get the digram or second order evaluation of closed phase in running speech.

Figure 14 DQx 1&2  distributions of first and second order “closed phase” 

as a function of vocal fold frequency, Fx

Voice quality is a complex attribute of voice but one important aspect comes from the regularity and duration of the closed phase from vocal fold cycle to cycle. First and second order plots can often give important information in regard to the physical nature of a pathological voice and here it is evident that speaker B has poor closed phase coherence.

Figure 15 “Closed phase” ratio Qx as a function of vocal fold frequency

In normal speech the variation of the closed phase at the level of discourse is a rule governed phenomenon whose rules differ from one person to another. Speaker A is a young woman and shows an effect of increasing breathiness with larynx frequency ( work by Evelyn Abberton). The pathological voice, B, is substantially deviant and gives a range of Qx [the closed phase measure based on trans-glottal conductance] which is never found in the normal voice and shows the irregularity as a function of frequency which can be clearly heard in her speaking voice.

Figure 16 A brief illustration of the use of precision timing in the simultaneous stroboscopic acquisition 

of vocal fold images and Lx/egg waveform information.

Figure 16 gives a simple example of the combination of the Lx/egg information that is used for the analysis of sustained and connected speech with stroboscopic images. The use of high precision timing to control the stroboscope is especially valuable in image acquisition. The immediate link to the Lx waveform is an aid to diagnosis. The system also allows the possibility of linking connected speech analyses to stroboscopic interpretation so that modes of abnormal vocal fold vibration can be better defined and investigated.

Acknowledgement

The analyses used have all been derived via the use of Laryngograph Ltd. Speech Studio and LxStrobe hardware and software.

1. References with particular relevance to pitch perception

[1] Nelson, D.A., Stanton, M. E., and Freyman R. L., 1983, “A general equation describing frequency discrimination as function of frequency and sensation level” J.A.S.A. 73, 2117-2123

[2] Vance, T.F.,1914, see page 340 in E.G. Boring, “Sensation and Perception in the History of Experimental Psychology”, Appleton-Century-Crofts Inc. New York 1942

[3] Wier, C.C., Jesteadt, W., Green, D.M., 1977, “Frequency discrimination as a function of frequency and sensation level”, J.A.S.A. 61, 178-184

[4] Moore, B.C.J. Glasberg, B.R., Peters, R.W., 1985, “Relative dominance of individual partials in determining the pitch of complex tones” J.A.S.A. 75, 550-561

[5] Klatt, D. H. 1973, “Discrimination of fundamental frequency contours in synthetic speech: implications for models of pitch perception”, J.A.S.A. 53, 8-16

[6] t’Haart, J.,1981, “Differential sensitivity to pitch distance, particularly in speech” J.A.S.A. 65, 811-821 Aronson, A. E. (1985) Clinical Voice Disorders. 2 nd edition Thieme Inc. NY Anastaplo, S. & Karnell, M.P. (1988) Synchronized videostroboscopic and electroglottographic examination of glottal opening. J.A.S.A 83, 1883-1890

2. References with particular relevance to voice evaluation

Baken, R. J., (1987) Clinical Measurement of Speech and Voice. Little, Brown & Company, Mass USA.

Cowie, R., Douglas-Cowie, E. and Rahilly, J. (1988) The intonation of adults with postlingually acquired deafness: anomalous frequency and distribution of elements. In B. Ainsworth and J. Holmes (Eds) Proceedings SPEECH '88, 7th FASE Symposium, Edinburgh, 2, 481-487.

Dejonckere, Ph., Remacle, M., Fresnel-Elbaz, E., Woisard, V., Crevier-Buchman, L., Millet, B. (1996) 

Differentiated perceptual evaluation of pathological voice quality: reliability and correlations with acoustic measurements. Rev. Laryngol. (Bordeaux) 117, 219-224 

Fabre,P., (1957) Un procédé électrique percutané d’inscription de l’accolement glottique au cours de la phonation. Bull. Nat. Méd., 141, 66-99 

Fourcin, A, & Norgate, M.(1965) Measurement of trans-glottal impedance. Progress Report, Phonetics laboratory University College London, pp 34-40 

Fourcin, A., Abberton, E. (1972) First applications of a new laryngograph Volta Review, 69, 507-518 {reprinted from Med. & Biol. Illustration 21, (1971), 172-182}.

Fourcin, A.J., (1974) Laryngographic examination of vocal fold vibration. in B.Wyke (ed.) Ventilatory and Phonatory Control Systems. OUP 1974, pp. 315-333 

Fourcin, A.J.,(1981) Laryngographic Assessment of Phonatory Function ASHA Reports #11 pp116-127 

Gilbert, H.R., Potter, C.R., & Hoodin, R (1984), Laryngograph as a measure of vocal fold contact area. J. Speech Hear Res., 27, 178-182 

Hammarberg, B. & Gauffin, J., (1995) Perceptual and acoustic characteristics of quality differences in pathological voices as related to physiological aspects,. in O. Fujimura & M.Hirano (eds.) Vocal Fold Physiology pp. 283-303 

Hirano, M., (1981) Clinical Examination of Voice. New York, Springer-Verlag 

Jones, C. (1967) 'Deaf Voice' - a description derived from a survey of the literature.Volta Review, 69, 507-8.

Kent, R.D., (1996) Hearing and Believing. Am. J. Sp-Lang Path., 5, 7-23 

Laukkanen, A.-M., & Vilkman, E. Tremor in the Light of Sound Production with Excised Human Larynges in P.H. Dejonckere, O.Hirano, J. Sunberg, (eds.) Vibrato 1995, Singular Publishing, pp. 93-110 

Laver, J. (1991), The Gift of Speech Edinburgh Un. Press 

Lecluse, F.L.E., (1977) Elektroglottografie. Drukerijelinkwijk B.V.

Maddieson, I., (1984) Patterns of Sounds, CUP Cambridge 

Scherer, R.C., Druker, D.G. & Titze, I.R. Electroglottography and Direct Measurement of Vocal Fold Contact Area. in O. Fujimura (ed.) vol.2, Vocal Fold Physiology, Raven Press, pp. 279-291 

Moore, G. P., (1971) Organic Voice Disorders. Englewood Cliffs, N.J., Prentice-Hall