Relative contributions of spectral and temporal cues for phoneme recognition

Cochlear implants provide users with limited spectral and temporal information. In this study, the amount of spectral and temporal information was systematically varied through simulations of cochlear implant processors using a noise-excited vocoder. Spectral information was controlled by varying the number of channels between 1 and 16, and temporal information was controlled by varying the lowpass cutoff frequencies of the envelope extractors from 1 to 512 Hz. Consonants and vowels processed using those conditions were presented to seven normal-hearing native-English-speaking listeners for identification. The results demonstrated that both spectral and temporal cues were important for consonant and vowel recognition with the spectral cues having a greater effect than the temporal cues for the ranges of numbers of channels and lowpass cutoff frequencies tested. The lowpass cutoff for asymptotic performance in consonant and vowel recognition was 16 and 4 Hz, respectively. The number of channels at which performance plateaued for consonants and vowels was 8 and 12, respectively. Within the above-mentioned ranges of lowpass cutoff frequency and number of channels, the temporal and spectral cues showed a tradeoff for phoneme recognition. Information transfer analyses showed different relative contributions of spectral and temporal cues in the perception of various phonetic/acoustic features.     

Speech intelligibility in cochlear implant simulations: Effects of carrier type, interfering noise, and subject experience

Channel vocoders using either tone or band-limited noise carriers have been used in experiments to simulate cochlear implant processing in normal-hearing listeners. Previous results from these experiments have suggested that the two vocoder types produce speech of nearly equal intelligibility in quiet conditions. The purpose of this study was to further compare the performance of tone and noise-band vocoders in both quiet and noisy listening conditions. In each of four experiments, normal-hearing subjects were better able to identify tone-vocoded sentences and vowel-consonant-vowel syllables than noise-vocoded sentences and syllables, both in quiet and in the presence of either speech-spectrum noise or two-talker babble. An analysis of consonant confusions for listening in both quiet and speech-spectrum noise revealed significantly different error patterns that were related to each vocoder's ability to produce tone or noise output that accurately reflected the consonant's manner of articulation. Subject experience was also shown to influence intelligibility. Simulations using a computational model of modulation detection suggest that the noise vocoder's disadvantage is in part due to the intrinsic temporal fluctuations of its carriers, which can interfere with temporal fluctuations that convey speech recognition cues.     

Speech recognition in noise as a function of the number of spectral channels: comparison of acoustic hearing and cochlear implants

Speech recognition was measured as a function of spectral resolution (number of spectral channels) and speech-to-noise ratio in normal-hearing (NH) and cochlear-implant (CI) listeners. Vowel, consonant, word, and sentence recognition were measured in five normal-hearing listeners, ten listeners with the Nucleus-22 cochlear implant, and nine listeners with the Advanced Bionics Clarion cochlear implant. Recognition was measured as a function of the number of spectral channels (noise bands or electrodes) at signal-to-noise ratios of + 15, + 10, +5, 0 dB, and in quiet. Performance with three different speech processing strategies (SPEAK, CIS, and SAS) was similar across all conditions, and improved as the number of electrodes increased (up to seven or eight) for all conditions. For all noise levels, vowel and consonant recognition with the SPEAK speech processor did not improve with more than seven electrodes, while for normal-hearing listeners, performance continued to increase up to at least 20 channels. Speech recognition on more difficult speech materials (word and sentence recognition) showed a marginally significant increase in Nucleus-22 listeners from seven to ten electrodes. The average implant score on all processing strategies was poorer than scores of NH listeners with similar processing. However, the best CI scores were similar to the normal-hearing scores for that condition (up to seven channels). CI listeners with the highest performance level increased in performance as the number of electrodes increased up to seven, while CI listeners with low levels of speech recognition did not increase in performance as the number of electrodes was increased beyond four. These results quantify the effect of number of spectral channels on speech recognition in noise and demonstrate that most CI subjects are not able to fully utilize the spectral information provided by the number of electrodes used in their implant.     

Relative contribution of off- and on-frequency spectral components of background noise to the masking of unprocessed and vocoded speech

The present study examined the relative influence of the off- and on-frequency spectral components of modulated and unmodulated maskers on consonant recognition. Stimuli were divided into 30 contiguous equivalent rectangular bandwidths. The temporal fine structure (TFS) in each "target" band was either left intact or replaced with tones using vocoder processing. Recognition scores for 10, 15 and 20 target bands randomly located in frequency were obtained in quiet and in the presence of all 30 masker bands, only the off-frequency masker bands, or only the on-frequency masker bands. The amount of masking produced by the on-frequency bands was generally comparable to that produced by the broadband masker. However, the difference between these two conditions was often significant, indicating an influence of the off-frequency masker bands, likely through modulation interference or spectral restoration. Although vocoder processing systematically lead to poorer consonant recognition scores, the deficit observed in noise could often be attributed to that observed in quiet. These data indicate that (i) speech recognition is affected by the off-frequency components of the background and (ii) the nature of the target TFS does not systematically affect speech recognition in noise, especially when energetic masking and/or the number of target bands is limited.     

Features of stimulation affecting tonal-speech perception: implications for cochlear prostheses

Tone languages differ from English in that the pitch pattern of a single-syllable word conveys lexical meaning. In the present study, dependence of tonal-speech perception on features of the stimulation was examined using an acoustic simulation of a CIS-type speech-processing strategy for cochlear prostheses. Contributions of spectral features of the speech signals were assessed by varying the number of filter bands, while contributions of temporal envelope features were assessed by varying the low-pass cutoff frequency used for extracting the amplitude envelopes. Ten normal-hearing native Mandarin Chinese speakers were tested. When the low-pass cutoff frequency was fixed at 512 Hz, consonant, vowel, and sentence recognition improved as a function of the number of channels and reached plateau at 4 to 6 channels. Subjective judgments of sound quality continued to improve as the number of channels increased to 12, the highest number tested. Tone recognition, i.e., recognition of the four Mandarin tone patterns, depended on both the number of channels and the low-pass cutoff frequency. The trade-off between the temporal and spectral cues for tone recognition indicates that temporal cues can compensate for diminished spectral cues for tone recognition and vice versa. An additional tone recognition experiment using syllables of equal duration showed a marked decrease in performance, indicating that duration cues contribute to tone recognition. A third experiment showed that recognition of processed FM patterns that mimic Mandarin tone patterns was poor when temporal envelope and duration cues were removed.     

Relative importance of temporal information in various frequency regions for consonant identification in quiet and in noise

The relative importance of temporal information in broad spectral regions for consonant identification was assessed in normal-hearing listeners. For the purpose of forcing listeners to use primarily temporal-envelope cues, speech sounds were spectrally degraded using four-noise-band vocoder processing Frequency-weighting functions were determined using two methods. The first method consisted of measuring the intelligibility of speech with a hole in the spectrum either in quiet or in noise. The second method consisted of correlating performance with the randomly and independently varied signal-to-noise ratio within each band. Results demonstrated that all bands contributed equally to consonant identification when presented in quiet. In noise, however, both methods indicated that listeners consistently placed relatively more weight upon the highest frequency band. It is proposed that the explanation for the difference in results between quiet and noise relates to the shape of the modulation spectra in adjacent frequency bands. Overall, the results suggest that normal-hearing listeners use a common listening strategy in a given condition. However, this strategy may be influenced by the competing sounds, and thus may vary according to the context. Some implications of the results for cochlear implantees and hearing-impaired listeners are discussed.     

Combined spectral and temporal enhancement to improve cochlear-implant speech perception

The present study examined the effect of combined spectral and temporal enhancement on speech recognition by cochlear-implant (CI) users in quiet and in noise. The spectral enhancement was achieved by expanding the short-term Fourier amplitudes in the input signal. Additionally, a variation of the Transient Emphasis Spectral Maxima (TESM) strategy was applied to enhance the short-duration consonant cues that are otherwise suppressed when processed with spectral expansion. Nine CI users were tested on phoneme recognition tasks and ten CI users were tested on sentence recognition tasks both in quiet and in steady, speech-spectrum-shaped noise. Vowel and consonant recognition in noise were significantly improved with spectral expansion combined with TESM. Sentence recognition improved with both spectral expansion and spectral expansion combined with TESM. The amount of improvement varied with individual CI users. Overall the present results suggest that customized processing is needed to optimize performance according to not only individual users but also listening conditions.     

Companding to improve cochlear-implant speech recognition in speech-shaped noise

Nonlinear sensory and neural processing mechanisms have been exploited to enhance spectral contrast for improvement of speech understanding in noise. The "companding" algorithm employs both two-tone suppression and adaptive gain mechanisms to achieve spectral enhancement. This study implemented a 50-channel companding strategy and evaluated its efficiency as a front-end noise suppression technique in cochlear implants. The key parameters were identified and evaluated to optimize the companding performance. Both normal-hearing (NH) listeners and cochlear-implant (CI) users performed phoneme and sentence recognition tests in quiet and in steady-state speech-shaped noise. Data from the NH listeners showed that for noise conditions, the implemented strategy improved vowel perception but not consonant and sentence perception. However, the CI users showed significant improvements in both phoneme and sentence perception in noise. Maximum average improvement for vowel recognition was 21.3 percentage points (p<0.05) at 0 dB signal-to-noise ratio (SNR), followed by 17.7 percentage points (p<0.05) at 5 dB SNR for sentence recognition and 12.1 percentage points (p<0.05) at 5 dB SNR for consonant recognition. While the observed results could be attributed to the enhanced spectral contrast, it is likely that the corresponding temporal changes caused by companding also played a significant role and should be addressed by future studies.     

Evaluation of two multichannel tactile aids for the hearing impaired

Two multichannel tactile devices for the hearing impaired were compared in speech perception tasks of varying levels of complexity. Both devices implemented the "vocoder" principle in their stimulus processing: One device had a 16-element linear vibratory array worn on the forearm and displayed activity in 16 overlapping frequency channels; the other device delivered tactile stimulation to a linear array of 16 electrodes worn on the abdomen. Subjects were tested in several phoneme discrimination tasks, ranging from discrimination of pairs of words differing in only one phoneme under tactile aid alone conditions to identification of stimuli in a larger set under tactile aid alone, lipreading alone, and lipreading plus tactile aid conditions. Results showed both devices to be better transmitters of manner and voicing features of articulation than of place features, when tested in single-item tasks. No systematic differences in performance with the two devices were observed. However, in a connected discourse tracking task, the vibrotactile vocoder in conjunction with lipreading yielded much greater improvements over lipreading alone than did the electrotactile vocoder. One possible explanation for this difference in performance, the inclusion of a noise suppression circuit in the electrotactile aid, was evaluated, but did not appear to account for the differences observed. Results are discussed in terms of additional differences between the two devices that may influence performance.     

Speech recognition with altered spectral distribution of envelope cues.

Recognition of consonants, vowels, and sentences was measured in conditions of reduced spectral resolution and distorted spectral distribution of temporal envelope cues. Speech materials were processed through four bandpass filters (analysis bands), half-wave rectified, and low-pass filtered to extract the temporal envelope from each band. The envelope from each speech band modulated a band-limited noise (carrier bands). Analysis and carrier bands were manipulated independently to alter the spectral distribution of envelope cues. Experiment I demonstrated that the location of the cutoff frequencies defining the bands was not a critical parameter for speech recognition, as long as the analysis and carrier bands were matched in frequency extent. Experiment II demonstrated a dramatic decrease in performance when the analysis and carrier bands did not match in frequency extent, which resulted in a warping of the spectral distribution of envelope cues. Experiment III demonstrated a large decrease in performance when the carrier bands were shifted in frequency, mimicking the basal position of electrodes in a cochlear implant. And experiment IV showed a relatively minor effect of the overlap in the noise carrier bands, simulating the overlap in neural populations responding to adjacent electrodes in a cochlear implant. Overall, these results show that, for four bands, the frequency alignment of the analysis bands and carrier bands is critical for good performance, while the exact frequency divisions and overlap in carrier bands are not as critical.     

Contribution of low-frequency acoustic information to Chinese speech recognition in cochlear implant simulations

Chinese sentence recognition strongly relates to the reception of tonal information. For cochlear implant (CI) users with residual acoustic hearing, tonal information may be enhanced by restoring low-frequency acoustic cues in the nonimplanted ear. The present study investigated the contribution of low-frequency acoustic information to Chinese speech recognition in Mandarin-speaking normal-hearing subjects listening to acoustic simulations of bilaterally combined electric and acoustic hearing. Subjects listened to a 6-channel CI simulation in one ear and low-pass filtered speech in the other ear. Chinese tone, phoneme, and sentence recognition were measured in steady-state, speech-shaped noise, as a function of the cutoff frequency for low-pass filtered speech. Results showed that low-frequency acoustic information below 500 Hz contributed most strongly to tone recognition, while low-frequency acoustic information above 500 Hz contributed most strongly to phoneme recognition. For Chinese sentences, speech reception thresholds (SRTs) improved with increasing amounts of low-frequency acoustic information, and significantly improved when low-frequency acoustic information above 500 Hz was preserved. SRTs were not significantly affected by the degree of spectral overlap between the CI simulation and low-pass filtered speech. These results suggest that, for CI patients with residual acoustic hearing, preserving low-frequency acoustic information can improve Chinese speech recognition in noise.     

Temporal envelope compensation for robust phoneme recognition using modulation spectrum

A robust feature extraction technique for phoneme recognition is proposed which is based on deriving modulation frequency components from the speech signal. The modulation frequency components are computed from syllable-length segments of sub-band temporal envelopes estimated using frequency domain linear prediction. Although the baseline features provide good performance in clean conditions, the performance degrades significantly in noisy conditions. In this paper, a technique for noise compensation is proposed where an estimate of the noise envelope is subtracted from the noisy speech envelope. The noise compensation technique suppresses the effect of additive noise in speech. The robustness of the proposed features is further enhanced by the gain normalization technique. The normalized temporal envelopes are compressed with static (logarithmic) and dynamic (adaptive loops) compression and are converted into modulation frequency features. These features are used in an automatic phoneme recognition task. Experiments are performed in mismatched train/test conditions where the test data are corrupted with various environmental distortions like telephone channel noise, additive noise, and room reverberation. Experiments are also performed on large amounts of real conversational telephone speech. In these experiments, the proposed features show substantial improvements in phoneme recognition rates compared to other speech analysis techniques. Furthermore, the contribution of various processing stages for robust speech signal representation is analyzed.     

Recognition of time-distorted sentences by normal-hearing and cochlear-implant listeners

This study evaluated the effects of time compression and expansion on sentence recognition by normal-hearing (NH) listeners and cochlear-implant (CI) recipients of the Nucleus-22 device. Sentence recognition was measured in five CI users using custom 4-channel continuous interleaved sampler (CIS) processors and five NH listeners using either 4-channel or 32-channel noise-band processors. For NH listeners, recognition was largely unaffected by time expansion, regardless of spectral resolution. However, recognition of time-compressed speech varied significantly with spectral resolution. When fine spectral resolution (32 channels) was available, speech recognition was unaffected even when the duration of sentences was shortened to 40% of their original length (equivalent to a mean duration of 40 ms/phoneme). However, a mean duration of 60 ms/phoneme was required to achieve the same level of recognition when only coarse spectral resolution (4 channels) was available. Recognition patterns were highly variable across CI listeners. The best CI listener performed as well as NH subjects listening to corresponding spectral conditions; however, three out of five CI listeners performed significantly poorer in recognizing time-compressed speech. Further investigation revealed that these three poorer-performing CI users also had more difficulty with simple temporal gap-detection tasks. The results indicate that limited spectral resolution reduces the ability to recognize time-compressed speech. Some CI listeners have more difficulty with time-compressed speech, as produced by rapid speakers, because of reduced spectral resolution and deficits in auditory temporal processing.     

Relative contribution of target and masker temporal fine structure to the unmasking of consonants in noise

The present study assessed the relative contribution of the "target" and "masker" temporal fine structure (TFS) when identifying consonants. Accordingly, the TFS of the target and that of the masker were manipulated simultaneously or independently. A 30 band vocoder was used to replace the original TFS of the stimuli with tones. Four masker types were used. They included a speech-shaped noise, a speech-shaped noise modulated by a speech envelope, a sentence, or a sentence played backward. When the TFS of the target and that of the masker were disrupted simultaneously, consonant recognition dropped significantly compared to the unprocessed condition for all masker types, except the speech-shaped noise. Disruption of only the target TFS led to a significant drop in performance with all masker types. In contrast, disruption of only the masker TFS had no effect on recognition. Overall, the present data are consistent with previous work showing that TFS information plays a significant role in speech recognition in noise, especially when the noise fluctuates over time. However, the present study indicates that listeners rely primarily on TFS information in the target and that the nature of the masker TFS has a very limited influence on the outcome of the unmasking process.     

Factors affecting the use of noise-band vocoders as acoustic models for pitch perception in cochlear implants

Although in a number of experiments noise-band vocoders have been shown to provide acoustic models for speech perception in cochlear implants (CI), the present study assesses in four experiments whether and under what limitations noise-band vocoders can be used as an acoustic model for pitch perception in CI. The first two experiments examine the effect of spectral smearing on simulated electrode discrimination and fundamental frequency (FO) discrimination. The third experiment assesses the effect of spectral mismatch in an FO-discrimination task with two different vocoders. The fourth experiment investigates the effect of amplitude compression on modulation rate discrimination. For each experiment, the results obtained from normal-hearing subjects presented with vocoded stimuli are compared to results obtained directly from CI recipients. The results show that place pitch sensitivity drops with increased spectral smearing and that place pitch cues for multi-channel stimuli can adequately be mimicked when the discriminability of adjacent channels is adjusted by varying the spectral slopes to match that of CI subjects. The results also indicate that temporal pitch sensitivity is limited for noise-band carriers with low center frequencies and that the absence of a compression function in the vocoder might alter the saliency of the temporal pitch cues.     

Binaural benefit for speech recognition with spectral mismatch across ears in simulated electric hearing

The present study investigated the effects of binaural spectral mismatch on binaural benefits in the context of bilateral cochlear implants using acoustic simulations. Binaural spectral mismatch was systematically manipulated by simulating changes in the relative insertion depths across ears. Sentence recognition, presented unilaterally and bilaterally, were measured in normal-hearing listeners in quiet and noise at +5 dB signal-to-noise ratio. Significant binaural benefits were observed when the interaural difference in insertion depth was 1 mm or less. This result suggests a dependence of the binaural benefit on redundant speech information, rather than on similarity in performance across ears.     

Effect of speech-intrinsic variations on human and automatic recognition of spoken phonemes

The aim of this study is to quantify the gap between the recognition performance of human listeners and an automatic speech recognition (ASR) system with special focus on intrinsic variations of speech, such as speaking rate and effort, altered pitch, and the presence of dialect and accent. Second, it is investigated if the most common ASR features contain all information required to recognize speech in noisy environments by using resynthesized ASR features in listening experiments. For the phoneme recognition task, the ASR system achieved the human performance level only when the signal-to-noise ratio (SNR) was increased by 15 dB, which is an estimate for the human-machine gap in terms of the SNR. The major part of this gap is attributed to the feature extraction stage, since human listeners achieve comparable recognition scores when the SNR difference between unaltered and resynthesized utterances is 10 dB. Intrinsic variabilities result in strong increases of error rates, both in human speech recognition (HSR) and ASR (with a relative increase of up to 120%). An analysis of phoneme duration and recognition rates indicates that human listeners are better able to identify temporal cues than the machine at low SNRs, which suggests incorporating information about the temporal dynamics of speech into ASR systems.     

Contribution of frequency modulation to speech recognition in noise

Cochlear implants allow most patients with profound deafness to successfully communicate under optimal listening conditions. However, the amplitude modulation (AM) information provided by most implants is not sufficient for speech recognition in realistic settings where noise is typically present. This study added slowly varying frequency modulation (FM) to the existing algorithm of an implant simulation and used competing sentences to evaluate FM contributions to speech recognition in noise. Potential FM advantage was evaluated as a function of the number of spectral bands, FM depth, FM rate, and FM band distribution. Barring floor and ceiling effects, significant improvement was observed for all bands from 1 to 32 with the additional FM cue both in quiet and noise. Performance also improved with greater FM depth and rate, which might reflect resolved sidebands under the FM condition. Having FM present in low-frequency bands was more beneficial than in high-frequency bands, and only half of the bands required the presence of FM, regardless of position, to achieve performance similar to when all bands had the FM cue. These results provide insight into the relative contributions of AM and FM to speech communication and the potential advantage of incorporating FM for cochlear implant signal processing.     

Evaluation of companding-based spectral enhancement using simulated cochlear-implant processing

This study tested a time-domain spectral enhancement algorithm that was recently proposed by Turicchia and Sarpeshkar [IEEE Trans. Speech Audio Proc. 13, 243-253 (2005)]. The algorithm uses a filter bank, with each filter channel comprising broadly tuned amplitude compression, followed by more narrowly tuned expansion (companding). Normal-hearing listeners were tested in their ability to recognize sentences processed through a noise-excited envelope vocoder that simulates aspects of cochlear-implant processing. The sentences were presented in a steady background noise at signal-to-noise ratios of 0, 3, and 6 dB and were either passed directly through an envelope vocoder, or were first processed by the companding algorithm. Using an eight-channel envelope vocoder, companding produced small but significant improvements in speech reception. Parametric variations of the companding algorithm showed that the improvement in intelligibility was robust to changes in filter tuning, whereas decreases in the time constants resulted in a decrease in intelligibility. Companding continued to provide a benefit when the number of vocoder frequency channels was increased to sixteen. When integrated within a sixteen-channel cochlear-implant simulator, companding also led to significant improvements in sentence recognition. Thus, companding may represent a readily implementable way to provide some speech recognition benefits to current cochlear-implant users.