Search

A system is described which uses a Honeywell DDP-516 computer to prepare and execute subjective evaluation tests of prerecorded speech material. The system utilizes two ordinary tape reproducers for the storage and fetching of speech samples. Both tape drives are under computer control. On-line access to any desired speech sample in the repertoire makes possible the use of sequential testing strategies. The capabilities and design limitations of the system are discussed. Two subjective investigations on the effects of the source function on speech quality have been carried out using this facility. These are described briefly.

Quantitative data on field strength and telephonic intelligibility are given for transmission at frequencies between 2.7 mega-cycles (111 m.) and 18 megacycles (16 m.), and for distances up to 1,000 miles with some data at 3,400 miles. The data are presented in the form of curves and surfaces, the variables being time of day, frequency and distance. Comparisons are made between transmission over land and over water, between night effects and day effects, and between transmission from horizontal and from vertical antennas. Fading, speech quality and noise are discussed. The results are briefly interpreted in terms of current short wave theories.

Starting from an IEEE Recommended Practice for Speech Quality Measurements and from previous work of the authors, a large set of vocoded speech signals has been evaluated in terms of preference. The set of speech samples has been taken from the vocoder survey of the 1967 Conference on Speech Communication and Processing, Boston, Mass. The test samples are evaluated by several methods: direct comparisons, the isopreferenee method, the relative preference method, the category judgment method, and the absolute preference judgment method. Due to the size of the test material, not all the test samples could be evaluated by all these methods. The test results are discussed and it is shown that, in certain respects, reliable system evaluations pose formidable problems. An effort to rank order the systems, which are described by small sets of test samples of frequently very different quality, for good reasons shows only limited success. The majority of the systems are of about equal preference with only insignificant differences. There are only a few systems that are outside this group and are either significantly better or worse than the rest.

Aperiodicity in voiced segments of speech may be ascribed to different causes. The magnitude of pitch perturbation is different in different spectral ranges of the signal. To see whether pitch perturbation can be effectively simulated by partially replacing voiced excitation by random noise, in appropriate frequency-time portions, experimental tests have been made on a computer-simulated channel vocoder. The buzz-hiss decision was made separately for three different frequency portions of the signal. The cepstrum technique was used for pitch detection, and separate buzz-hiss switching decisions were made at the synthesizer for each frequency portion. The switching thresholds were controlled, and deliberately "devoiced" versions were compared with regular vocoded speech. The fundamental frequency was determined by the lowband cepstrum. The result shows that partial devoicing of the high-frequency ranges definitely improves speech quality. Further, a comparatively large amount of devoicing is perceptually tolerable.

The objectives of this investigation were twofold: 1) to demonstrate the utility of the nine-point speech "goodness" rating scale as a method for scaling user opinion of speech quality, and 2) to use this method to determine optimum parameters for differential pulse-code modulation (DPCM) systems with bit rates from 25.6 to 51.2 kb/s. Fifteen DPCM and pulse-code modulation (PCM) systems were simulated on a digital computer. The parameters investigated included the tradeoff between bandwidth and number of quantization levels, and the number of taps in the DPCM predictor network. A total of 248 ratings were obtained from 31 trained listeners for each of the systems under consideration. Both the intra- and interrater reliability of these data, as obtained from the speech "goodness" rating scale, were found to be greater than 0.95. Results indicated that: 1) at any bit rate, DPCM is significantly better than PCM; 2)DPCM with a three-tap predictor is not significantly better than DPCM with a one-tap predictor; 3)between 2.4 and 4.3 kHz changes in bandwidth are inconsequential in terms of user opinion; and 4) the number of quantization bits appears to be the primary determinant of speech quality judgment.

Much time and study is being devoted to the improvement of Vocoder speech quality and intelligibility. One factor which is being neglected, however, is the effect of transmission errors on the intelligibility of Vocoded speech. To evaluate this effect, a full duplex 16-Channel Philco Vocoder is employed. Errors are inserted into the PCM output of the transmitter at various rates and in a random fashion. Intelligibility tests are run to determine the loss of intelligibility with increasing error rates. Curves are plotted showing these effects for a number of speakers. Fixed position errors are also inserted within the pitch portion of the bit frame and similar intelligibility tests are run. Results for the random insertion of errors indicate that an error rate of one to two per cent (1 to 2 bits per 100) is tolerable before any appreciable loss in the zero error PB word intelligibility of 82 per cent is apparent. At a transmission error rate approaching five per cent (5 incorrect bits per 100), the Channel Vocoder is capable of a PB word intelligibility of 70 per cent, corresponding to a sentence intelligibility of 90 per cent. Comments on the loss of speech quality with increasing errors are also given. Error insertion in the pitch portion of the frame indicates a surprising ability of the Channel Vocoder to tolerate error while maintaining intelligibility.

The use of the presurgical maxillary prosthesis in patients undergoing maxillectomy procedures for cancer greatly decreases the general morbidity connected with the procedure and aids the patient to adjust both physiologically and psychologically following the operation. Excellent speech quality and oral function can be attained almost universally. The historical, surgical, and prosthetic considerations concerning the presurgical maxillary prosthesis have been discussed, and the treatment of a patient is described to demonstrate the surgical and prosthetic principles involved in the management of cancer of the maxilla.

In 170 laryngectomized the radiologic profile pictures of the hypopharynx during deglutition were compared with X-ray pictures during phonation of the vowel ‘a’. Removing or leaving the hyoid bone does neither affect the incidence of recess formation northeir size. In most cases a cicatricial, transversal fold was found. It ran below the recess in the frontal plane and was very much alike an epiglottis. There was a statistically significant correlation between the substitute speech quality and the behaviour of hypopharynx recesses during phonation. In the patients with pharyngeal speech and whisper, the number of deep recesses was higher than that of shallow ones.

In some communications systems, the need arises for temporally interpolating data or signalling information during continuous speech.1 If the required time gaps are created by simply interrupting the speech signal, severe degradation of speech quality and some loss in intelligibility results.

Tape recordings made in the helium‐oxygen atmosphere of the U. S. Navy's SEALAB II have been analyzed by spectrographic techniques. The divers breathing this gas mixture have an unusual voice quality often characterized as “Donald Duck” speech. The Sealab experiment permitted speech analysis of divers living in such an atmosphere for several days. A preliminary study of the recorded data has led to the following observations: (1) formant shifts are responsible for the unusual quality of the helium‐oxygen speech; (2) the formant shifts are nonlinear—the frequency shift of the first formant is greater than the frequency shifts of the higher order formants; (3) relative formant amplitudes are often changed radically; (4) pitch or fundamental frequency changes are not significant; (5) after several days in this new atmosphere, changes occurred in the speech quality that made it sound more natural. These changes most likely have been affected by modifications of the vocal‐tract gestures.

Consider one voice, bandlimited in K different ways. A complete paired‐comparison schedule is presented to N listeners to obtain “similarity‐with‐respect‐to‐quality” judgments. From each listener's judgments, a “perceptual space” is determined and the question is: “Are there individual differences between these N perceptual spaces?” In order to answer this, we compare perceptual spaces, obtain some index of similarity, and determine a “points‐of‐view” space containing N points. Suppose this is a two‐space. Typically we now argue for differences on the basis of how the N points spread out on a sheet of graph paper, but a very real question is rarely considered. How big can the points‐of‐view space be? By introducing “random listeners” (random perceptual spaces), we can answer this question; and the results so far indicate that “real listeners” cluster so tightly in this “real‐plus‐random” space that no individual differences can be found for the “real listeners.” [This research was supported by the Office of Naval Research.]

Although there is widespread use of category scales for the evaluation of speech quality, the labeling and instructions to subjects about the use of such scales has frequently been intuitively selected. In this experiment, subjects rated speech stimuli varying in volume and circuit noise level on a verbal (“Unsatisfactory”‐“Excellent”) or a numerical (1–5) scale. Subjects using the verbal scale were instructed either to define the endpoints of the scale as equal to the highest and lowest signal/noise ratios in the experiment (external anchors) or to use their telephony experience to define the scale (internal anchors). The subjects using the numerical scale used external anchors only. On the verbal scale, the internal anchoring resulted in a higher mean rating and more peaked response distribution than the external anchoring; this indicates that subjects can apply this scale somewhat independently of the stimulus distribution. In comparing the verbal and numerical scales, the response distributions show that there are small differences in category boundaries between the two scales. In general, the results suggest that anchoring plays a more important role than scale labeling in determining the subjects’ response distributions.

The familiar model of the speech spectrum as a product of the source spectrum times the vocal‐tract transfer function suggests making an actual separation of the two for two different talkers followed by a recombining of the vocal tract of one talker with the source of the other. The calculations to effect such an interchange were made on a pitch‐synchronous basis. Fourier coefficients were obtained for each period, the harmonic‐amplitude spectrum was matched to obtain locations of the vocal tract poles, and these poles were divided from the spectrum. The pole data for an utterance by one talker could thus be recombined with the source spectral data for an utterance by another talker and the speech wave regenerated by Fourier synthesis. Source data for several talkers were tried, each with the same vocal tract and also a single source was applied to several “heads.” Listening tests were made of the resulting combinations in an attempt to determine the effect of glottal source versus vocal tract on speech quality and talker identification.

The interference test for speech quality evaluation was investigated for sensitivity and validity with various types of telephone speech. Three components of the telephone system (microphone, transmission channel, and earphone), each with two alternatives (high fidelity or telephone), were used in all possible combinations to obtain eight types of telephone speech. The interference test was sensitive enough to measure the amount by which each telephone component degraded speech, even though all the speech types were highly intelligible according to articulation tests. Consistent and systematic relationships were observed between the interference test results, articulation test scores, and subjective evaluations of the same speech types, indicating that the interference test is a valid procedure for measuring speech quality. Moreover, by simply adding up the interference test measures associated with the telephone components, the measure for a speech type with any combination of the components could be acc...

The process of listening to speech over earphones, and simultaneously repeating it is called “shadowing.” The use of shadowing as a possible method for evaluating speech transmission quality was investigated. Three methods were used for degrading speech quality: additive noise, speech correlated noise, and peak clipping. All speech was presented over earphones. Subjects practiced the shadowing test to an 85% corrcet criterion on undegraded speech, with a word presentation rate of 1.31 words/sec. A male and a female speaker were used. With all three degradations, shadowing accuracy was monotonically increasing with decreasing degradation. For a given signal‐to‐noise ratio, performance was better while listening to speech correlated noise than while listening to additive noise. Shadowing was found to be more sensitive to low levels of peak clipping than most types of intelligibility tests.

The need often exists to screen for hearing acuity a group of subjects that are to be used in subsequent experiments on auditory discrimination or speech transmission quality evaluation. A method has been developed by which the existing multiple listening facility of a laboratory principally devoted to speech quality measurement was adapted to enable screening audiograms to be given to as many as 22 subjects simultaneously. Subjects listen to the stimuli, which are groups of up to three tone pulses, over modified telephone handsets. Subjects are instructed to respond positively only if they hear a specified number of pulses. This has the effect of forcing a conservative criterion upon the subjects, which reduces response variability. Tests are made at seven frequencies between 250 Hz and 4 kHz. Tone levels are varied from 30 dB hearing level to −5 dB hearing level (re ISO‐1964) in 5 dB steps. The subject responds on a keyboard that is connected to an SDS‐930 computer. This computer entirely controls the t...

Naturally uttered sentences were analyzed and then resynthesized with the range of pitch and formant excursions varied systematically; these sentences were ranked for preference, clarity, and naturalness in separate sessions. The ranges of the synthesis parameters f0, f1, f2, f3 were compressed or expanded about their respective geometric means in the log frequency domain by a factor k; pitch (k=0–1.60) and formant (k=0.55–1.20, all formants transformed simultaneously) variations were factorially combined to generate the sentence ensemble. The data were analyzed using multidimensional scaling procedures that accounted for differences among the listeners. The analysis showed that preference was related to both pitch and formants; clarity was generally related to formants; and naturalness was idiosyncratic. Formant compression or expansion generally lowered both preference and clarity rankings. The influence of pitch compression or expansion on preference varied with the range of the formant excursion.

A new method for evaluating speech quality was investigated to supplement articulation tests. Subjects recalled a single target syllable from a randomized list presented at various rates. The target syllable either immediately preceded or followed a predesignated cue syllable which occurred at random serial positions. Results: (1) recall deteriorated as rate increased; (2) this deterioration was more pronounced when the target preceded the cue; and (3) serial position had a negligible effect on recall. These results suggest that, when the target precedes the cue, subjects attempts to recognize and encode into memory each syllable before the next one occurs. Encoding, which is essential for recall and which apparently cannot proceed in parallel with recognition, may be interrupted when the next syllable arrives and requires recognition. Such interruptions reduce recall and occur more often as recognition time increases. If recognition time increases as speech quality declines, then the recall level in the target‐cue condition should be a sensitive measure of speech quality. Apparently subjects do not alternate between recognition and encoding in the cue‐target condition, thus precluding the interruptions from which recognition times are inferred.

This letter describes the design of a differential far-end-operated halfecho suppressor to be used on telephone circuits having long propagation times. Emphasis has been placed on reducing the deterioration of the speech quality, caused by residual echo and speech chopping.

In a previous paper in 1963, a method was described for the measurement of speech quality. It was proposed to utilize a reference signal consisting of high‐quality speech degraded by additive white noise. Extensive studies have shown disadvantages of this form of degradation. The reference signal differs considerably in its nature from the output speech signals generated by most speech‐processing devices, and in spite of its apparently simple definition it cannot be reproduced very easily, because of the difficulties in measuring the sound level of speech signals. A new reference signal is proposed. Instead of adding a controlled amount of noise to a high‐quality speech signal, the noise is multiplied into the signal. One way to generate the reference signal of the form s(t)[1+k⋅r(t)] is to modify this form to s(t)+k⋅r(t)⋅s(t) and to generate the product with a Hall multiplier. Test results are given and discussed. [Research reported in this paper sponsored by the U. S. Government.]

An experiment was designed to study the effects of speech and noise levels on speech quality using white and speech‐correlated noise (SCN). Two types of speech signals were used: (1) flat with respect to frequency (flat speech), and (2) flat below 1 kHz, and with frequency emphasis of 9 dB/oct above 1 kHz (shaped speech). Both speech and noise were bandlimited between 150 and 3600 Hz. The speech level ranged between 54 and 100 dBt SPL and the noise spectrum level between 2 and 33 dBt/Hz. A total of 45 noise and speech level combinations were used as test stimuli. Each test condition consisted of a pair of phonetically balanced sentences in the presence of one of the 45 test stimuli. The subjects were required to rate the speech quality on a 5‐point scale. Averaged over all conditions, the shaped speech was preferred over flat speech. For shaped speech, SCN was preferred over white noise. For flat speech the same was true for speech levels of about 80 dBt and higher, and the reverse was true for speech lev...