Classification of stop place in consonant-vowel contexts using feature extrapolation of acoustic-phonetic features in telephone speech

Knowledge-based speech recognition systems extract acoustic cues from the signal to identify speech characteristics. For channel-deteriorated telephone speech, acoustic cues, especially those for stop consonant place, are expected to be degraded or absent. To investigate the use of knowledge-based methods in degraded environments, feature extrapolation of acoustic-phonetic features based on Gaussian mixture models is examined. This process is applied to a stop place detection module that uses burst release and vowel onset cues for consonant-vowel tokens of English. Results show that classification performance is enhanced in telephone channel-degraded speech, with extrapolated acoustic-phonetic features reaching or exceeding performance using estimated Mel-frequency cepstral coefficients (MFCCs). Results also show acoustic-phonetic features may be combined with MFCCs for best performance, suggesting these features provide information complementary to MFCCs.     

A probabilistic framework for landmark detection based on phonetic features for automatic speech recognition

A probabilistic framework for a landmark-based approach to speech recognition is presented for obtaining multiple landmark sequences in continuous speech. The landmark detection module uses as input acoustic parameters (APs) that capture the acoustic correlates of some of the manner-based phonetic features. The landmarks include stop bursts, vowel onsets, syllabic peaks and dips, fricative onsets and offsets, and sonorant consonant onsets and offsets. Binary classifiers of the manner phonetic features-syllabic, sonorant and continuant-are used for probabilistic detection of these landmarks. The probabilistic framework exploits two properties of the acoustic cues of phonetic features-(1) sufficiency of acoustic cues of a phonetic feature for a probabilistic decision on that feature and (2) invariance of the acoustic cues of a phonetic feature with respect to other phonetic features. Probabilistic landmark sequences are constrained using manner class pronunciation models for isolated word recognition with known vocabulary. The performance of the system is compared with (1) the same probabilistic system but with mel-frequency cepstral coefficients (MFCCs), (2) a hidden Markov model (HMM) based system using APs and (3) a HMM based system using MFCCs.     

Evaluation of formant-like features on an automatic vowel classification task

Numerous attempts have been made to find low-dimensional, formant-related representations of speech signals that are suitable for automatic speech recognition. However, it is often not known how these features behave in comparison with true formants. The purpose of this study was to compare two sets of automatically extracted formant-like features, i.e., robust formants and HMM2 features, to hand-labeled formants. The robust formant features were derived by means of the split Levinson algorithm while the HMM2 features correspond to the frequency segmentation of speech signals obtained by two-dimensional hidden Markov models. Mel-frequency cepstral coefficients (MFCCs) were also included in the investigation as an example of state-of-the-art automatic speech recognition features. The feature sets were compared in terms of their performance on a vowel classification task. The speech data and hand-labeled formants that were used in this study are a subset of the American English vowels database presented in Hillenbrand et al. [J. Acoust. Soc. Am. 97, 3099-3111 (1995)]. Classification performance was measured on the original, clean data and in noisy acoustic conditions. When using clean data, the classification performance of the formant-like features compared very well to the performance of the hand-labeled formants in a gender-dependent experiment, but was inferior to the hand-labeled formants in a gender-independent experiment. The results that were obtained in noisy acoustic conditions indicated that the formant-like features used in this study are not inherently noise robust. For clean and noisy data as well as for the gender-dependent and gender-independent experiments the MFCCs achieved the same or superior results as the formant features, but at the price of a much higher feature dimensionality.     

Automatic classification and speaker identification of African elephant (Loxodonta africana) vocalizations

A hidden Markov model (HMM) system is presented for automatically classifying African elephant vocalizations. The development of the system is motivated by successful models from human speech analysis and recognition. Classification features include frequency-shifted Mel-frequency cepstral coefficients (MFCCs) and log energy, spectrally motivated features which are commonly used in human speech processing. Experiments, including vocalization type classification and speaker identification, are performed on vocalizations collected from captive elephants in a naturalistic environment. The system classified vocalizations with accuracies of 94.3% and 82.5% for type classification and speaker identification classification experiments, respectively. Classification accuracy, statistical significance tests on the model parameters, and qualitative analysis support the effectiveness and robustness of this approach for vocalization analysis in nonhuman species.     

Predicting fundamental frequency from mel-frequency cepstral coefficients to enable speech reconstruction

This work proposes a method to reconstruct an acoustic speech signal solely from a stream of mel-frequency cepstral coefficients (MFCCs) as may be encountered in a distributed speech recognition (DSR) system. Previous methods for speech reconstruction have required, in addition to the MFCC vectors, fundamental frequency and voicing components. In this work the voicing classification and fundamental frequency are predicted from the MFCC vectors themselves using two maximum a posteriori (MAP) methods. The first method enables fundamental frequency prediction by modeling the joint density of MFCCs and fundamental frequency using a single Gaussian mixture model (GMM). The second scheme uses a set of hidden Markov models (HMMs) to link together a set of state-dependent GMMs, which enables a more localized modeling of the joint density of MFCCs and fundamental frequency. Experimental results on speaker-independent male and female speech show that accurate voicing classification and fundamental frequency prediction is attained when compared to hand-corrected reference fundamental frequency measurements. The use of the predicted fundamental frequency and voicing for speech reconstruction is shown to give very similar speech quality to that obtained using the reference fundamental frequency and voicing.     

The relative phonetic contributions of a cochlear implant and residual acoustic hearing to bimodal speech perception

The addition of low-passed (LP) speech or even a tone following the fundamental frequency (F0) of speech has been shown to benefit speech recognition for cochlear implant (CI) users with residual acoustic hearing. The mechanisms underlying this benefit are still unclear. In this study, eight bimodal subjects (CI users with acoustic hearing in the non-implanted ear) and eight simulated bimodal subjects (using vocoded and LP speech) were tested on vowel and consonant recognition to determine the relative contributions of acoustic and phonetic cues, including F0, to the bimodal benefit. Several listening conditions were tested (CI/Vocoder, LP, T(F0-env), CI/Vocoder + LP, CI/Vocoder + T(F0-env)). Compared with CI/Vocoder performance, LP significantly enhanced both consonant and vowel perception, whereas a tone following the F0 contour of target speech and modulated with an amplitude envelope of the maximum frequency of the F0 contour (T(F0-env)) enhanced only consonant perception. Information transfer analysis revealed a dual mechanism in the bimodal benefit: The tone representing F0 provided voicing and manner information, whereas LP provided additional manner, place, and vowel formant information. The data in actual bimodal subjects also showed that the degree of the bimodal benefit depended on the cutoff and slope of residual acoustic hearing.     

Toward a model for lexical access based on acoustic landmarks and distinctive features

This article describes a model in which the acoustic speech signal is processed to yield a discrete representation of the speech stream in terms of a sequence of segments, each of which is described by a set (or bundle) of binary distinctive features. These distinctive features specify the phonemic contrasts that are used in the language, such that a change in the value of a feature can potentially generate a new word. This model is a part of a more general model that derives a word sequence from this feature representation, the words being represented in a lexicon by sequences of feature bundles. The processing of the signal proceeds in three steps: (1) Detection of peaks, valleys, and discontinuities in particular frequency ranges of the signal leads to identification of acoustic landmarks. The type of landmark provides evidence for a subset of distinctive features called articulator-free features (e.g., [vowel], [consonant], [continuant]). (2) Acoustic parameters are derived from the signal near the landmarks to provide evidence for the actions of particular articulators, and acoustic cues are extracted by sampling selected attributes of these parameters in these regions. The selection of cues that are extracted depends on the type of landmark and on the environment in which it occurs. (3) The cues obtained in step (2) are combined, taking context into account, to provide estimates of "articulator-bound" features associated with each landmark (e.g., [lips], [high], [nasal]). These articulator-bound features, combined with the articulator-free features in (1), constitute the sequence of feature bundles that forms the output of the model. Examples of cues that are used, and justification for this selection, are given, as well as examples of the process of inferring the underlying features for a segment when there is variability in the signal due to enhancement gestures (recruited by a speaker to make a contrast more salient) or due to overlap of gestures from neighboring segments.     

Speech waveform envelope cues for consonant recognition

This study investigated the cues for consonant recognition that are available in the time-intensity envelope of speech. Twelve normal-hearing subjects listened to three sets of spectrally identical noise stimuli created by multiplying noise with the speech envelopes of 19(aCa) natural-speech nonsense syllables. The speech envelope for each of the three noise conditions was derived using a different low-pass filter cutoff (20, 200, and 2000 Hz). Average consonant identification performance was above chance for the three noise conditions and improved significantly with the increase in envelope bandwidth from 20-200 Hz. SINDSCAL multidimensional scaling analysis of the consonant confusions data identified three speech envelope features that divided the 19 consonants into four envelope feature groups ("envemes"). The enveme groups in combination with visually distinctive speech feature groupings ("visemes") can distinguish most of the 19 consonants. These results suggest that near-perfect consonant identification performance could be attained by subjects who receive only enveme and viseme information and no spectral information.     

The ability of cochlear implant users to use temporal envelope cues recovered from speech frequency modulation

Previous studies have demonstrated that normal-hearing listeners can understand speech using the recovered "temporal envelopes," i.e., amplitude modulation (AM) cues from frequency modulation (FM). This study evaluated this mechanism in cochlear implant (CI) users for consonant identification. Stimuli containing only FM cues were created using 1, 2, 4, and 8-band FM-vocoders to determine if consonant identification performance would improve as the recovered AM cues become more available. A consistent improvement was observed as the band number decreased from 8 to 1, supporting the hypothesis that (1) the CI sound processor generates recovered AM cues from broadband FM, and (2) CI users can use the recovered AM cues to recognize speech. The correlation between the intact and the recovered AM components at the output of the sound processor was also generally higher when the band number was low, supporting the consonant identification results. Moreover, CI subjects who were better at using recovered AM cues from broadband FM cues showed better identification performance with intact (unprocessed) speech stimuli. This suggests that speech perception performance variability in CI users may be partly caused by differences in their ability to use AM cues recovered from FM speech cues.     

Recognition of natural and time/intensity altered CVs by young and elderly subjects with normal hearing

The purpose of this study was to evaluate the efficiency of three acoustic modifications derived from clear speech for improving consonant recognition by young and elderly normal-hearing subjects. Percent-correct nonsense syllable recognition was measured for four stimulus sets: unmodified stimuli; stimuli with consonant duration increased by 100%; stimuli with consonant-vowel ratio increased by 10 dB; and stimuli with both consonant duration and consonant-vowel ratio increased. Analyses of overall nonsense syllable recognition, consonant feature recognition, and consonant confusion patterns demonstrated that the consonant-vowel ratio increase modification produced better performance than the other acoustic modifications by both subject groups. However, elderly subjects exhibited poorer performance than young subjects in most conditions. These results and their implications are discussed.