Gammatone Filterbank Research Articles

Sub band Speech analysis using Gammatone Filter banks and optimal pitch extraction methods for each sub band using average magnitude difference function (AMDF) for LPC Speech Coders in Noisy Environments

Modern speech processing applications require operation on signal of interest that is contaminated by high level of noise. These situations call for a greater robustness in estimation of the speech parameters for mismatch environment and low environmental SNR level. In this paper the speech is analyzed with a Gammatone filter bank. This splits the full band speech signal s(n) into frequency bands(sub bands).and for each sub band speech signal pitch is extracted. We determine the Signal to Noise Ratio for each Sub band speech signal. Then the average of pitch periods of the highest SNR sub bands is used to obtain a optimal pitch value. This paper describes a computationally simple Pitch extraction algorithms using Average Magnitude Difference Function (AMDF) which is a new approach using weighted Autocorrelation 2 and very useful for accurate Pitch Period extraction. Both these algorithms can be software implemented and performance evaluated. Both of them uses center clipping for time domain processing. This paper also in general Compares the effectiveness of the new AMDF using weighted Autocorrelation and the existing Autocorrelation method and how it is possible to utilize this further in Speech Enhancement Systems using the proposed new algorithms for its implementation.

Noise-robust acoustic signature recognition using nonlinear Hebbian learning

We propose using a new biologically inspired approach, nonlinear Hebbian learning (NHL), to implement acoustic signal recognition in noisy environments. The proposed learning processes both spectral and temporal features of input acoustic data. The spectral analysis is realized by using auditory gammatone filterbanks. The temporal dynamics is addressed by analyzing gammatone-filtered feature vectors over multiple temporal frames, which is called a spectro-temporal representation (STR). Given STR features, the exact acoustic signatures of signals of interest and the mixing property between signals of interest and noises are generally unknown. The nonlinear Hebbian learning is then employed to extract representative independent features from STRs, and to reduce their dimensionality. The extracted independent features of signals of interest are called signatures. In the meantime of learning, the synaptic weight vectors between input and output neurons are adaptively updated. These weight vectors project data into a feature subspace, in which signals of interest are selected, while noises are attenuated. Compared with linear Hebbian learning (LHL) which explores the second-order moment of data, the applied NHL involves the higher-order statistics of data. Therefore, NHL can capture representative features that are more statistically independent than LHL can. Besides, the nonlinear activation function of NHL can be chosen to refer to the implicit distribution of many acoustic sounds, and thus making the learning optimized in an aspect of mutual information. Simulation results show that the whole proposed system can more accurately recognize signals of interest than other conventional methods in severely noisy circumstances. One applicable project is detecting moving vehicles. Noise-contaminated vehicle sound is recognized while other non-vehicle sounds are rejected. When vehicle is contaminated by human vowel, bird chirp, or additive white Gaussian noise (AWGN) at SNR=0dB, the proposed system dramatically decreases the error rate over normally used acoustic feature extraction method, mel-frequency cepstral computation (MFCC), by 26%, 36.3%, and 60.3%, respectively; and, over LHL by 20%, 2.3%, and 15.3%, respectively. Another applicable project is vehicle type identification. The proposed system achieves better performance than LHL, e.g., 40% improvement when gasoline heavy wheeled car is contaminated by AWGN at SNR=5dB. More importantly, the proposed system is implemented in real-time field testing for months. The purpose is to detect vehicle with any make or model moving on the street with speed 10-35mph. The missing rate is 1-2%, when vehicle is contaminated by any surrounding noises (human conversation, animal sound, airplane, wind, etc.) at SNR=0-20dB. The false alarm rate is around 1%. To summarize, this study not only provides an efficient approach to extract representative independent features from high-dimensional data, but also offers robustness against severe noises.

Acoustic features for speech recognition based on Gammatone filterbank and instantaneous frequency

Most of the features used by modern automatic speech recognition systems, such as mel-frequency cepstral coefficients (MFCC) and perceptual linear predictive (PLP) coefficients, represent spectral envelope of the speech signal only. Nevertheless, phase or frequency modulation as represented in recent perceptual models of the peripheral auditory system might also contribute to speech decoding. Furthermore, such features can be complementary to the envelope features. This paper proposes a variety of features based on a linear auditory filterbank, the Gammatone filterbank. Envelope features are derived from the envelope of the subband filter outputs. Phase/frequency modulation is represented by the subband instantaneous frequency (IF) and is used explicitly by concatenating envelope-based and IF-based features or is used implicitly by IF-based frequency reassignment. Speech recognition experiments using a standard HMM-based recognizer under both clean training and multi-condition training are conducted on a Chinese mandarin digits corpus. The experimental results show that the proposed envelope and phase based features can improve recognition rates in clean and noisy conditions compared to the reference MFCC-based recognizer.

Combining Auditory Preprocessing and Bayesian Estimation for Robust Formant Tracking

We present a framework for estimating formant trajectories. Its focus is to achieve high robustness in noisy environments. Our approach combines a preprocessing based on functional principles of the human auditory system and a probabilistic tracking scheme. For enhancing the formant structure in spectrograms we use a Gammatone filterbank, a spectral preemphasis, as well as a spectral filtering using difference-of-Gaussians (DoG) operators. Finally, a contrast enhancement mimicking a competition between filter responses is applied. The probabilistic tracking scheme adopts the mixture modeling technique for estimating the joint distribution of formants. In conjunction with an algorithm for adaptive frequency range segmentation as well as Bayesian smoothing an efficient framework for estimating formant trajectories is derived. Comprehensive evaluations of our method on the VTR-formant database emphasize its high precision and robustness. We obtained superior performance compared to existing approaches for clean as well as echoic noisy speech. Finally, an implementation of the framework within the scope of an online system using instantaneous feature-based resynthesis demonstrates its applicability to real-world scenarios.

Auditory-model-based Feature Extraction Method for Mechanical Faults Diagnosis

It is well known that the human auditory system possesses remarkable capabilities to analyze and identify signals. Therefore, it would be significant to build an auditory model based on the mechanism of human auditory systems, which may improve the effects of mechanical signal analysis and enrich the methods of mechanical faults features extraction. However the existing methods are all based on explicit senses of mathematics or physics, and have some shortages on distinguishing different faults, stability, and suppressing the disturbance noise, etc. For the purpose of improving the performances of the work of feature extraction, an auditory model, early auditory(EA) model, is introduced for the first time. This auditory model transforms time domain signal into auditory spectrum via bandpass filtering, nonlinear compressing, and lateral inhibiting by simulating the principle of the human auditory system. The EA model is developed with the Gammatone filterbank as the basilar membrane. According to the characteristics of vibration signals, a method is proposed for determining the parameter of inner hair cells model of EA model. The performance of EA model is evaluated through experiments on four rotor faults, including misalignment, rotor-to-stator rubbing, oil film whirl, and pedestal looseness. The results show that the auditory spectrum, output of EA model, can effectively distinguish different faults with satisfactory stability and has the ability to suppress the disturbance noise. Then, it is feasible to apply auditory model, as a new method, to the feature extraction for mechanical faults diagnosis with effect.

Voice biometric feature using Gammatone filterbank and ICA

Voice biometric feature extraction is the core task in developing any speaker identification system. This paper proposes a robust feature extraction technique for the purpose of speaker identification. The technique is based on processing monaural speech signal using human auditory system based Gammatone Filterbank (GTF) and Independent Component Analysis (ICA). The measures used to assess the robustness to additive noises and speaker identification performance are defined and discussed. The kkn the proposed feature is evaluated in real environments under varying noisy conditions. The proposed feature is benchmarked against the commonly used features such as: MFCC, PLCC, and PLP, and it outperforms them in different noisy environments.

Modulation Spectral Features for Robust Far-Field Speaker Identification

In this paper, auditory inspired modulation spectral features are used to improve automatic speaker identification (ASI) performance in the presence of room reverberation. The modulation spectral signal representation is obtained by first filtering the speech signal with a 23-channel gammatone filterbank. An eight-channel modulation filterbank is then applied to the temporal envelope of each gammatone filter output. Features are extracted from modulation frequency bands ranging from 3-15 H z and are shown to be robust to mismatch between training and testing conditions and to increasing reverberation levels. To demonstrate the gains obtained with the proposed features, experiments are performed with clean speech, artificially generated reverberant speech, and reverberant speech recorded in a meeting room. Simulation results show that a Gaussian mixture model based ASI system, trained on the proposed features, consistently outperforms a baseline system trained on mel-frequency cepstral coefficients. For multimicrophone ASI applications, three multichannel score combination and adaptive channel selection techniques are investigated and shown to further improve ASI performance.

Analysis and design of gammatone signal models

An established model for the signal analysis performed by the human cochlea is the overcomplete gammatone filterbank. The high correlation of this signal model with human speech and environmental sounds [E. Smith and M. Lewicki, Nature (London) 439, 978-982 (2006)], combined with the increased time-frequency resolution of sparse overcomplete signal models, makes the overcomplete gammatone signal model favorable for signal processing applications on natural sounds. In this paper a signal-theoretic analysis of overcomplete gammatone signal models using the theory of frames and performing bifrequency analyses is given. For the number of gammatone filters M> or =100 (2.4 filters per equivalent rectangular bandwidth), a near-perfect reconstruction can be achieved for the signal space of natural sounds. For signal processing applications like multi-rate coding, a signal-to-alias ratio can be used to derive decimation factors with minimal aliasing distortions.

The role of spectral- and temporal-envelope room-acoustic cues in auditory selective attention.

Listeners can attend to one of several simultaneous messages by tracking one speaker’s voice characteristics. Using differences in the location of sounds in a room, we ask how well cues arising from spatial position compete with these characteristics. Listeners decided which of two simultaneous target words belonged in an attended “context” phrase when it was played simultaneously with a different “distracter” context. Talker difference was in competition with position difference, so the response indicates which cue-type the listener was tracking. Spatial position was found to override talker difference in dichotic conditions when the talkers are similar (male). The salience of cues associated with differences in sounds, bearings decreased with distance between listener and sources. These cues are more effective binaurally. However, there appear to be other cues that increase in salience with distance between sounds. This increase is more prominent in diotic conditions, indicating that these cues are largely monaural. Distances between spectra calculated using a gammatone filterbank (with ERB-spaced CFs) of the room’s impulse responses at different locations were computed, and comparison with listeners’ responses suggested some slight monaural loudness cues, but also monaural “timbre” cues arising from the temporal- and spectral-envelope differences in the speech from different locations.

Extending the sawtooth wave form inspired pitch estimator with an auditory-based preprocessing stage.

The sawtooth wave form inspired pitch estimator (SWIPE) [A. Camacho, J. Acoust. Soc. Am. 124, 1638–1652 (2008)] is extended by introducing an auditory-based preprocessing stage consisting of two main steps. First, a system consisting of a gammatone filterbank followed by half-wave rectifiers and low-pass filters is applied to the signal. This system tends to boost low-order harmonics of the signal, as shown in a work by Klapuri [A. Klapuri, IEEE Trans. Audio, Speech, Lang. Process. 16, 255–266 (2008)]. Second, the square root of the amplitude spectrum is computed on every channel, and they are summed to form a summary square-root spectrum (SSRS). SWIPE is then applied to this SSRS. Results show that the proposed method significantly outperforms the original SWIPE on music and telephonelike speech, albeit producing a small decrease in performance on high-fidelity speech.

Simulating the precedence effect by means of autocorrelation.

A number of algorithms have been developed to simulate the robust localization performance of humans in reverberant conditions, the most successful of which are based on (contralateral) inhibition. While these models can demonstrate the precedence effect in general, their nonlinear behavior makes it difficult to optimize their settings. A linear algorithm has now been developed to overcome this limitation. An autocorrelation algorithm determines the delay between lead and lag and their amplitude ratio for both channels. An inverse filter is then used to eliminate the lag signal before it is localized with a standard localization algorithm. Interestingly, the filter contains both inhibitory and excitatory elements, and the filter’s impulse response looks somewhat similar to the response of a chopper cell. The algorithm operates robustly on top of a model of the auditory periphery (gammatone filterbank and halfwave rectification). Due to its linear nature, the model performs better if the full waveform is reconstructed by subtracting a delayed version of the halfwave‐rectified signal, with a delay time that corresponds to half the period of each frequency band’s center frequency. The model is able to simulate a number of experiments with ongoing stimuli and performs robustly with onset‐truncated and interaural‐level‐difference based stimuli.

The simulation of bat-oriented auditory processing using the experimental data of echolocating signals

There are various approaches to understanding the echolocation phenomenon of bats. A part of the echolocating process is assessed here by determining what acoustic signal a bat's ears receive during echolocation. It is simplified in an experimental rig to measure the reflections from objects in different horizontal angles which represents a sound localisation task in bats. It has been assumed in this study that the remarkable echolocating ability of bats, which is not shown in the most other mammalian species, is achieved by their different physical shape of head and ears, and specialised auditory processing of echolocating signals. In human studies in terms of sound localisation, physical characteristics are usually modelled as a head-related transfer function (or HRTF) and gammatone filter banks are widely used to simulate auditory processing in the cochlear. A modified filterbank is used here to represent the auditory processing in bats and combined with the experimental data of object reflections. Bat HRTFs will be used subsequently to determine the acoustic reflections at both ears.

Across-frequency processes involved in auditory detection of coloration

When an early wall reflection is added to a direct sound, a spectral modulation is introduced to the signal's power spectrum. This spectral modulation typically produces an auditory sensation of coloration or pitch. Throughout this study, auditory spectral-integration effects involved in coloration detection are investigated. Coloration detection thresholds were therefore measured as a function of reflection delay and stimulus bandwidth. In order to investigate the involved auditory mechanisms, an auditory model was employed that was conceptually similar to the peripheral weighting model [Yost, JASA, 1982, 416-425]. When a “classical” gammatone filterbank was applied within this spectrum-based model, the model largely underestimated human performance at high signal frequencies. However, this limitation could be resolved by employing an auditory filterbank with narrower filters. This novel filterbank was designed to approximate auditory filter-shapes measured by Oxenham and Shera [JARO, 2003, 541-554], derived from forward masking data. The results of the present study demonstrate that a “purely” spectrum-based model approach can successfully describe auditory coloration detection even at high signal frequencies.

Amplitude modulation shape and speech intelligibility

Research has demonstrated that low frequency amplitude modulations in speech signals are crucially important to maintaining intelligibility. The current work demonstrates a flexible way of characterising the pulsatile bursts of energy found in the temporal envelopes of sub‐band filtered speech. Speech was passed through a 128‐filter Gammatone filterbank and the temporal envelope of each filter extracted using the Hilbert transform. Thirty‐five raised cosine pulses were fitted to model the envelope of each filter and each pulse was defined by its amplitude, half‐duration and centre position. The distribution of these pulses demonstrates that the most commonly found pulse half‐duration in speech is around 10 ms and few pulses have half‐durations longer than 25 ms. Highly intelligible vocoded speech is generated using the extracted pulses and these measures suggest that the auditory system may signal the position in time of the amplitude modulations rather than representing low‐frequency information. This me...

Dynamic compression in hearing aids based on an auditory model

A multichannel dynamic compression algorithm is proposed that uses a novel nonlinear auditory filterbank which aims at effectively describing the basilar membrane (BM) response to arbitrary signals. It is based on a linear Gammatone filterbank, a subsequent instantaneous compression stage and a frequency-synthesis stage. In order to model the linear response to off-frequency tones and suppression effects from signals below the characteristic frequency (low-side suppression) the compression characteristics is controlled in each filter band by the deviation of the current sub-band instantaneous frequency from the band's center frequency. If the deviation is small, on-frequency components are prominent, and full gain and compression is applied. If the deviation is large, off-frequency components are prominent, and gain and compression is reduced. Simulations of responses to sinusoids as a function of frequency and level at a fixed BM place (i.e., within a fixed filter band) and for all filter bands (i.e., BM excitation patterns) show good correspondence with psychoacoustical excitation pattern models. Responses to two-tone stimuli quantitatively simulate psychoacoustical two-tone suppression. Likewise, nonlinear growth of simultaneous masking is quantitatively modeled. First results show that speech reception in modulated noise may be improved by the system for a subgroup of hearing-impaired subjects.

Representing multiple discrimination cues in a computational model of the bottlenose dolphin auditory system

A computational model of the dolphin auditory system was developed to describe how multiple discrimination cues may be represented and employed during echolocation discrimination tasks. The model consisted of a bank of gammatone filters followed by half-wave rectification and low pass filtering. The output of the model resembles a spectrogram; however, the model reflects temporal and spectral resolving properties of the dolphin auditory system. Model outputs were organized to represent discrimination cues related to spectral, temporal and intensity information. Two empirical experiments, a phase discrimination experiment [Johnson et al., Animal Sonar Processes and Performance (Plenum, New York, 1988)] and a cylinder wall thickness discrimination tasks [Au and Pawolski, J. Comp. Physiol. A 170, 41-47 (1992)] were then simulated. Model performance was compared to dolphin performance. Although multiple discrimination cues were potentially available to the dolphin, simulation results suggest temporal information was used in the former experiment and spectral information in the latter. This model's representation of sound provides a more accurate approximation to what the dolphin may be hearing compared to conventional spectrograms, time-amplitude, or spectral representations.

Prediction of speech intelligibility in spatial noise and reverberation for normal-hearing and hearing-impaired listeners

Binaural speech intelligibility of individual listeners under realistic conditions was predicted using a model consisting of a gammatone filter bank, an independent equalization-cancellation (EC) process in each frequency band, a gammatone resynthesis, and the speech intelligibility index (SII). Hearing loss was simulated by adding uncorrelated masking noises (according to the pure-tone audiogram) to the ear channels. Speech intelligibility measurements were carried out with 8 normal-hearing and 15 hearing-impaired listeners, collecting speech reception threshold (SRT) data for three different room acoustic conditions (anechoic, office room, cafeteria hall) and eight directions of a single noise source (speech in front). Artificial EC processing errors derived from binaural masking level difference data using pure tones were incorporated into the model. Except for an adjustment of the SII-to-intelligibility mapping function, no model parameter was fitted to the SRT data of this study. The overall correlation coefficient between predicted and observed SRTs was 0.95. The dependence of the SRT of an individual listener on the noise direction and on room acoustics was predicted with a median correlation coefficient of 0.91. The effect of individual hearing impairment was predicted with a median correlation coefficient of 0.95. However, for mild hearing losses the release from masking was overestimated.

Application of psychoacoustic filtering for machine fault detection

A method for nondestructive machine fault detection based on the evaluation of acoustic machine signatures is proposed. Various mechanical defects of rotary machines can be reflected in altered acoustic signatures. The proposed method is based on the psychoacoustic modelling of human auditory perception. For the purpose of machine fault detection, the gammatone filterbank is applied in the preprocessing of acoustic signals. Filtered signals are then rectified and features are calculated as mean values of rectified signals. A set of features represents an extracted machine signature. The machine states are evaluated by comparing the extracted signature with the database of previously collected machine signatures. The proposed approach is illustrated by a case study where the quality of commercially produced compressors is estimated. Results show that major faults that occur in a production can be reliably detected.

Effects of Instantaneous Multiband Dynamic Compression on Speech Intelligibility

The recruitment phenomenon, that is, the reduced dynamic range between threshold and uncomfortable level, is attributed to the loss of instantaneous dynamic compression on the basilar membrane. Despite this, hearing aids commonly use slow-acting dynamic compression for its compensation, because this was found to be the most successful strategy in terms of speech quality and intelligibility rehabilitation. Former attempts to use fast-acting compression gave ambiguous results, raising the question as to whether auditory-based recruitment compensation by instantaneous compression is in principle applicable in hearing aids. This study thus investigates instantaneous multiband dynamic compression based on an auditory filterbank. Instantaneous envelope compression is performed in each frequency band of a gammatone filterbank, which provides a combination of time and frequency resolution comparable to the normal healthy cochlea. The gain characteristics used for dynamic compression are deduced from categorical loudness scaling. In speech intelligibility tests, the instantaneous dynamic compression scheme was compared against a linear amplification scheme, which used the same filterbank for frequency analysis, but employed constant gain factors that restored the sound level for medium perceived loudness in each frequency band. In subjective comparisons, five of nine subjects preferred the linear amplification scheme and would not accept the instantaneous dynamic compression in hearing aids. Four of nine subjects did not perceive any quality differences. A sentence intelligibility test in noise (Oldenburg sentence test) showed little to no negative effects of the instantaneous dynamic compression, compared to linear amplification. A word intelligibility test in quiet (one-syllable rhyme test) showed that the subjects benefit from the larger amplification at low levels provided by instantaneous dynamic compression. Further analysis showed that the increase in intelligibility resulting from a gain provided by instantaneous compression is as high as from a gain provided by linear amplification. No negative effects of the distortions introduced by the instantaneous compression scheme in terms of speech recognition are observed.

Hearing relative phases for two harmonic components

There is a new wavelet model of basilar membrane motion which predicts that existing roex and gammatone filterbanks underestimate the interaction of harmonically related components in complex tones. This interaction appears as a modulation of the basilar membrane motion associated with a higher harmonic by the presence of a lower harmonic; the period of the modulation is that of the difference frequency, or the fundamental of the implied harmonic series. The modulation depth is largest for stimuli whose spectra have a similar energy distribution as natural sounds: a high frequency roll-off of about 12-dB per octave. The strength of the modulation frequency is dependent on the relative phase of the components, which predicts that listeners will be able to hear a relative phase change. A series of experiments were undertaken to determine the threshold for a change in the relative phase of the components. The results of the experiments show that a change in the relative phase is detectable for components that are generally believed to be resolved on the basilar membrane and therefore should not interact. [Work supported by the U.K. Medical Research Council (G9901257).]