When Spectral Smearing Can Increase Speech Intelligibility.

Sentences were reduced to an array of sixteen effectively rectangular bands (RBs) having center frequencies ranging from 0.25 to 8 kHz spaced at ⅓-octave intervals. Four arrays were employed, each having uniform subcritical bandwidths which ranged from 40 Hz to 5 Hz. The 40 Hz width array had intelligibility near ceiling, and the 5 Hz array about 1%. The finding of interest was that when the subcritical speech RBs were used to modulate RBs of noise having the same center frequency as the speech but having bandwidths increased to a critical (ERBn) bandwidth at each center frequency, these spectrally smeared arrays were considerably more intelligible in all but the 40 Hz (ceiling) condition. For example, when the 10 Hz bandwidth speech array having an intelligibility of 8% modulated the ERBn noise array, intelligibility increased to 48%. This six-fold increase occurred despite elimination of spectral fine structure and addition of stochastic fluctuation to speech envelope cues. (As anticipated, conventional vocoding with matching bandwidths of speech and noise reduced the 10-Hz-speech array intelligibility from 8% to 1%). These effects of smearing confirm findings by Bashford, Warren, and Lenz (2010) that optimal temporal processing requires stimulation of a critical bandwidth. [Supported by NIH].

How Broadband Speech May Avoid Neural Firing Rate Saturation at High Intensities and Maintain Intelligibility.

While broadband speech may remain perfectly intelligible at levels exceeding 90 dB, narrowband speech intelligibility (e.g., 2/3-octave passband centered at 1.5 kHz) may decline by 25% or more at moderate intensities (e.g., 75 dB). This "rollover" effect is substantially reduced, however, when a speech band is accompanied by flanking bands of white noise [J.A. Bashford, R.M. Warren, & P.W. Lenz, 2005, J. Acoust. Soc. Am. 117, 365-369 (2005)], suggesting that lateral suppression helps preserve broadband speech intelligibility at high levels. The present study found that when noise flankers were presented individually at a low spectrum level (-30 dB relative to the speech) only the higher-frequency flanker produced a significant intelligibility increase. However, the lower-frequency flanking noise did produce an equivalent increase when its spectrum level was raised 10 dB. This asymmetrical intensity requirement for noise flankers links the effective dynamic range of speech intelligibility to reported characteristics of both lateral (two-tone) suppression of auditory nerve (AN) fiber activity and lateral inhibition of secondary cells of the cochlear nucleus. These and other observations will be discussed in the broader context of how various auditory mechanisms help preserve speech intelligibility at high intensities by reducing firing rate saturation. [Supported by NIH.].

Detection of acoustic repetition for very long stochastic patterns.

Guttman and Julesz (1963) employed recycling frozen noise segments (RFNs) as model stimuli in their classic study of the lower limits for periodicity detection and short-term auditory memory. They reported that listeners can hear iteration of these stochastic signals effortlessly as "motorboating" for repetition periods ranging from 50 to 250 msec and as "whooshing" from 250 msec to 1 sec. Both motorboating and whooshing RFNs are global percepts encompassing the entire period, as are RFNs in the pitch range (repetition periods shorter than 50 msec). However, with continued listening to whooshing (but not motorboating) RFNs, individuals hear recurrent brief components such as clanks and thumps that are characteristic of the particular waveform. Experiment 1 of the present study describes a cross-modal cuing procedure that enables listeners to store and then recognize the recurrence of portions of frozen noise waveforms that are repeated after intervals of 10 sec or more. Experiment 2 compares the relative saliencies of different spectral regions in enabling listeners to detect repetition of these long-period patterns. Special difficulty was encountered with the 6-kHz band of RFNs, possibly due to the lack of fine-structure phase locking at this frequency range. In addition, a similarity is noted between the organizational principles operating over particular durational ranges of stochastic patterns and the characteristics of traditional hierarchical units of speech having corresponding durations.

Intelligibility of bandpass speech: effects of truncation or removal of transition bands.

An intelligibility of over 90% was reported for keywords in "everyday" 1/3-octave sentences centered on 1500 Hz and having steep transition band slopes of 100 dB/octave [Warren et al., Percept. Psychophys. 57, 175-182 (1995)]. A subsequent study by Warren and Bashford [J. Acoust. Soc. Am. 106, L47-L52 (1999)] found that it was not the 1/3-octave passband, but the transition bands that were chiefly responsible for this high intelligibility: When the passband and transition bands were segregated using filter slopes of 1000 dB/octave, the isolated passband had an intelligibility score of only 24%, while the pair of transition bands had a score of over 80%. In the present study, experiment 1 examined the distribution of information along the transition bands' slopes by truncation at graded downpoints: Truncation at downpoints of 40 dB or more produced no significant change in intelligibility. Experiment 2 closed the gap separating the transition bands so that their slopes intersected at 1500 Hz. This triangular band had a negligible passband (as defined conventionally by 3-dB downpoints) and an intelligibility score of 60%; truncation at downpoints of 50 dB or more produced no significant change in intelligibility. Experiment 3 determined the intelligibilities of rectangular bands (1000-dB/octave slopes) centered on 1500 Hz. Their bandwidths ranged from 3 to 12 semitones in 1-semitone steps, resulting in intelligibility scores increasing monotonically from 14% to 94%. Calculations based upon experiments 2 and 3 showed that the triangular band truncated at 30-dB downpoints had half the intelligibility of a rectangular band having the same frequency range.

Intelligibility of 1/3-octave speech: greater contribution of frequencies outside than inside the nominal passband.

We reported previously that "everyday" sentences were highly intelligible when limited to a 1/3-octave passband centered at 1,500 Hz and having transition-band slopes of approximately 100 dB/octave. The present study determined the relative contributions to intelligibility made by the passband (PB) and the transition bands (TBs) by partitioning the same bandpass sentences using 2,000-order FIR filtering. Intelligibility scores were: PB with both TBs, 92%; deletion of both TBs (leaving only the 1/3-octave PB with nearly vertical slopes), 24%; deletion of the PB (leaving both TBs separated by a 1/3-octave gap), 83%. These and other results indicate a remarkable ability to compensate for severe spectral tilt and the consequent importance of considering frequencies outside the nominal passband in interpreting studies using filtered speech.

Spectral restoration of speech: intelligibility is increased by inserting noise in spectral gaps.

In order to function effectively as a means of communication, speech must be intelligible under the noisy conditions encountered in everyday life. Two types of perceptual synthesis have been reported that can reduce or cancel the effects of masking by extraneous sounds: Phonemic restoration can enhance intelligibility when segments are replaced or masked by noise, and contralateral induction can prevent mislateralization by effectively restoring speech masked at one ear when it is heard in the other. The present study reports a third type of perceptual synthesis induced by noise: enhancement of intelligibility produced by adding noise to spectral gaps. In most of the experiments, the speech stimuli consisted of two widely separated narrow bands of speech (center frequencies of 370 and 6,000 Hz, each band having high-pass and low-pass slopes of 115 dB/octave meeting at the center frequency). These very narrow bands effectively reduced the available information to frequency-limited patterns of amplitude fluctuation lacking information concerning formant structure and frequency transitions. When stochastic noise was introduced into the gap separating the two speech bands, intelligibility increased for "everyday" sentences, for sentences that varied in the transitional probability of keywords, and for monosyllabic word lists. Effects produced by systematically varying noise amplitude and noise bandwidth are reported, and the implications of some of the novel effects observed are discussed.

Use of speech-modulated noise adds strong "bottom-up" cues for phonemic restoration.

When deleted segments of speech are replaced by extraneous sounds rather than silence, the missing speech fragments may be perceptually restored and intelligibility improved. This phonemic restoration (PhR) effect has been used to measure various aspects of speech processing, with deleted portions of speech typically being replaced by stochastic noise. However, several recent studies of PhR have used speech-modulated noise, which may provide amplitude-envelope cues concerning the replaced speech. The present study compared the effects upon intelligibility of replacing regularly spaced portions of speech with stochastic (white) noise versus speech-modulated noise. In Experiment 1, filling periodic gaps in sentences with noise modulated by the amplitude envelope of the deleted speech fragments produced twice the intelligibility increase obtained with interpolated stochastic noise. Moreover, when lists of isolated monosyllables were interrupted in Experiment 2, interpolation of speech-modulated noise increased intelligibility whereas stochastic noise reduced intelligibility. The augmentation of PhR produced by modulated noise appeared without practice, suggesting that speech processing normally involves not only a narrowband analysis of spectral information but also a wideband integration of amplitude levels across critical bands. This is of considerable theoretical interest, but it also suggests that since PhRs produced by speech-modulated noise utilize potent bottom-up cues provided by the noise, they differ from the PhRs produced by extraneous sounds, such as coughs and stochastic noise.

Spectral redundancy: intelligibility of sentences heard through narrow spectral slits.

The intelligibility of word lists subjected to various types of spectral filtering has been studied extensively. Although words used for communication are usually present in sentences rather than lists, there has been no systematic report of the intelligibility of lexical components of narrowband sentences. In the present study, we found that surprisingly little spectral information is required to identify component words when sentences are heard through narrow spectral slits. Four hundred twenty listeners (21 groups of 20 subjects) were each presented with 100 bandpass filtered CID ("everyday speech") sentences; separate groups received center frequencies of 370, 530, 750, 1100, 1500, 2100, 3000, 4200, and 6000 Hz at 70 dBA SPL. In Experiment 1, intelligibility of single 1/3-octave bands with steep filter slopes (96 dB/octave) averaged more than 95% for sentences centered at 1100, 1500, and 2100 Hz. In Experiment 2, we used the same center frequencies with extremely narrow bands (slopes of 115 dB/octave intersecting at the center frequency, resulting in a nominal bandwidth of 1/20 octave). Despite the severe spectral tilt for all frequencies of this impoverished spectrum, intelligibility remained relatively high for most bands, with the greatest intelligibility (77%) at 1500 Hz. In Experiments 1 and 2, the bands centered at 370 and 6000 Hz provided little useful information when presented individually, but in each experiment they interacted synergistically when combined. The present findings demonstrate the adaptive flexibility of mechanisms used for speech perception and are discussed in the context of the LAME model of opportunistic multilevel processing.

Auditory induction: reciprocal changes in alternating sounds.

When portions of a sound are replaced by a potential masker, the missing fragments may be perceptually restored, resulting in apparent continuity of the interrupted signal. This phenomenon has been examined extensively by using pulsation threshold, auditory induction, and phonemic restoration paradigms in which two sounds, the inducer and the inducee, are alternated (ABABA...), and the conditions required for apparent continuity of the lower amplitude inducee are determined. Previous studies have generally neglected to examine concomitant changes produced in the inducing sound. Results from the present experiments have demonstrated decreases in the loudness of inducers using inducer/inducee pairs consisting of tone/tone and noise/noise, as well as the noise/speech pairs associated with phonemic restorations. Interestingly, reductions in inducer loudness occurred even when the inducee was heard as discontinuous, and these decreases in loudness were accompanied by graded increases in apparent duration of the inducee, contrary to the conventional view of auditory induction as an all-or-none phenomenon. Under some conditions, the reduced loudness of the inducer was coupled with a marked alteration in its timbre. Especially profound changes in the inducer quality occurred when the alternating stimuli were tones having the same frequency and differing only in intensity--it seems that following subtraction of components corresponding to the inducee, an anomalous auditory residue remained that did not correspond to the representation of a tone.

When acoustic sequences are not perceptual sequences: the global perception of auditory patterns.

Warren, Bashford, and Gardner (1990) found that when sequences consisting of 10 40-msec steady-state vowels were presented in recycled format, minimal changes in order (interchanging the position of two adjacent phonemes) produced easily recognizable differences in verbal organization, even though the vowel durations were well below the threshold for identification of order. The present study was designed to determine if this ability to discriminate between different arrangements of components is limited to speech sounds subject to verbal organization, or if it reflects a more general auditory ability. In the first experiment, 10 40-msec sinusoidal tones were substituted for the vowels; it was found that the easy discrimination of minimal changes in order is not limited to speech sounds. A second experiment substituted 10 40-msec frozen noise segments for the vowels. The succession of noise segments formed a 400-msec frozen noise pattern that cannot be considered as a sequence of individual sounds, as can the succession of vowels or tones. Nevertheless, listeners again could discriminate between patterns differing only in the order of two adjacent 40-msec segments. These results, together with other evidence, indicate that it is not necessary for acoustic sequences of brief items (such as phonemes and tones) to be processed as perceptual sequences (that is, as a succession of discrete identifiable sounds) for different arrangements to be discriminated. Instead, component acoustic elements form distinctive "temporal compounds," which permit listeners to distinguish between different arrangements of portions of an acoustic pattern without the need for segmentation into an ordered series of component items.(ABSTRACT TRUNCATED AT 250 WORDS)

Increasing the intelligibility of speech through multiple phonemic restorations.

Outside of the laboratory, listening conditions are often less than ideal, and when attending to sounds from a particular source, portions are often obliterated by extraneous noises. However, listeners possess rather elegant reconstructive mechanisms. Restoration can be complete, so that missing segments are indistinguishable from those actually present and the listener is unaware that the signal is fragmented. This phenomenon, called temporal induction (TI), has been studied extensively with nonverbal signals and to a lesser extent with speech. Earlier studies have demonstrated that TI can produce illusory continuity spanning gaps of a few hundred milliseconds when portions of a signal are replaced by a louder sound capable of masking the signal were it actually present. The present study employed various types of speech signals with periodic gaps and measured the effects upon intelligibility produced by filling these gaps with noises. Enhancement of intelligibility through multiple phonemic restoration occurred when the acoustic requirements for TI were met and when sufficient contextual information was available in the remaining speech fragments. It appears that phonemic restoration is a specialized form of TI that uses linguistic skills for the reconstruction of obliterated speech.

Tweaking the lexicon: organization of vowel sequences into words.

The ability of listeners to distinguish between different arrangements of the same three vowels was investigated for repeating sequences having item durations ranging from 10 msec (single glottal pulses) up to several seconds/vowel. Discrimination was accomplished with ease by untrained subjects at all item durations. From 30 through 100 msec/vowel, an especially interesting phenomenon was encountered: The sequences of steady-state vowels were organized into words, with different words heard for the different arrangements of items. In a second experiment, repeating sequences of random arrangements of 10 40-msec vowels were employed. When sets of four such sequences were presented to listeners, distinctive words were heard, which permitted each arrangement to be discriminated from the others. In addition, minimal differences (reversing the order of a single contiguous pair of vowels) in the 10-item sequences could be detected via verbal mediation. Hypotheses are offered concerning mechanisms responsible for these results.

Perception of complex tone pairs mistuned from unison.

Periodic sounds mistuned from unison may interact to produce pitch glides: When a broad-spectrum complex tone having a fundamental frequency of 400 Hz or less and containing several harmonics above the 8th is mixed with itself after a slight change in the waveform repetition frequency (1 Hz or less), listeners hear a rising glissando when corresponding portions of the waveforms approach alignment and a falling glissando as they recede from alignment. Glissandi are unimpaired if harmonics below the 8th are absent, but if, instead, harmonics above the 8th are removed, only amplitude fluctuations are heard (not glissandi). When two broad-spectrum complex tones with independent, randomly derived phase spectra are mistuned slightly from unison and mixed, complex repeated patterns other than glissandi are heard. These observations, along with others involving a variety of periodic sounds mistuned from unison, provide information concerning the nature of frequency domain and time domain mechanisms employed for the perception of iterated acoustic patterns.

Illusory continuity of interrupted speech: speech rate determines durational limits.

Deleted segments of speech can be restored perceptually if they are replaced by a louder noise. An earlier study of this "phonemic restoration effect" found that, when recorded discourse was interrupted periodically by noise, the durational limit for illusory continuity corresponded to the average word duration. The present study employed a different passage of discourse recorded by a different speaker. Durational limits for apparent continuity of discourse interrupted by noise were measured at the normal (original) playback speed, as well as at rates that were 15% greater and 15% less. At the normal playback rate, once again the limit of continuity approximated the average word duration--but of especial interest was the finding that changes in playback rate produced proportional changes in continuity limits. These results, together with other evidence, suggest that phonemic restorations represent a special linguistic application of a general auditory mechanism (auditory induction) producing appropriate syntheses of obliterated sounds, and that for discourse the limits of illusory continuity correspond to a fixed amount of verbal information, and not a fixed temporal value.