A feature extraction method using subband based periodicity and aperiodicity decomposition with noise robust frontend processing for automatic speech recognition

Ishizuka, Kentaro; Nakatani, Tomohiro

doi:10.1016/j.specom.2006.06.008

Cited by 17 publications

(7 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, the ability to hear tones in noise has important implications for speech intelligibility under noisy conditions (Plomp, 1994). In support of this, the detection of common periodicity has been successfully used to segment parts of noisy speech signals that contain vowel sounds (Hu and Wang, 2008), and the use of periodicity-based features was found to enhance the automatic speech recognition rates of voiced components of speech in the presence of noise (Ishizuka & Nakatani, 2006). Other researchers have considered the role of periodicity processing in segregating vowels by autocorrelation and similar algorithms such as recurrent neural networks (Assmann and Summerfield, 1990;Cariani, 2001;de Cheveigné and Kawahara, 1999;Meddis and Hewitt, 1992), and in the formation of stable auditory images of periodic sounds through the cross-channel temporal alignment of the maxima of auditory nerve spike rates (Patterson et al, 1995).…”

Section: Introductionmentioning

confidence: 93%

Enhancement of speech perception in noise by periodicity processing: A neurobiological model and signal processing algorithm

McLachlan

Grayden

2014

Speech Communication

View full text Add to dashboard Cite

The perceived loudness of sound increases with its tonality or periodicity, and the pitch strength of tones are linearly proportional to their sound pressure level. These observations suggest a fundamental relationship between pitch strength and loudness. This relationship may be explained by the superimposition of inputs to inferior colliculus neurons from cochlear nucleus chopper cells and phase locked spike trains from the lateral lemniscus. The regularity of chopper cell outputs increases for stimuli with periodicity at the same frequency as their intrinsic chopping rate. So inputs to inferior colliculus cells become synchronized for periodic stimuli, leading to increased likelihood that they will fire and increased salience of periodic signal components at the characteristic frequency of the inferior colliculus cell. A computer algorithm to enhance speech in noise was based on this model. The periodicity of the outputs of a Gammatone filter bank after each sound onset was determined by first sampling each filter channel at a range of typical chopper cell frequencies and then passing these amplitudes through a step function to simulate the firing of coincidence detecting neurons in the inferior colliculus. Filter channel amplification was based on the maximum accumulated spike count after each onset, resulting in increased amplitudes for filter channels with greater periodicity. The speech intelligibility of stimuli in noise was not changed when the algorithm was used to remove around 14 dB of noise from stimuli with signal-noise ratios of around 0 dB. This mechanism is a likely candidate for enhancing speech recognition in noise, and raises the proposition that pitch itself is an epiphenomenon that evolved from neural mechanisms that boost the hearing sensitivity of animals to vocalizations.

show abstract

Section: Introductionmentioning

confidence: 93%

Enhancement of speech perception in noise by periodicity processing: A neurobiological model and signal processing algorithm

McLachlan

Grayden

2014

Speech Communication

View full text Add to dashboard Cite

show abstract

“…This principle was first explored successfully in the architecture of deep auto encoder on the "raw" spectrogram or linear filter-bank features, [56] showing its superiority over the Mel-Cepstral features which contain a few stages of fixed transformation from spectrograms. The true "raw" features of speech, waveforms, have more recently been shown to produce excellent larger-scale speech recognition results [57].…”

Section: -2017smentioning

confidence: 99%

Enhancing Single Speaker Recognition Using Deep Belief Network

Prasetio

Hayashida

Nishizaki

et al. 2018

TMLAI

View full text Add to dashboard Cite

Recognition in speech is complex phenomena study, and the reason for this is the complexity of human language. The barrier of the problem in speech recognition study now can be handled from speech signal using machine learning methods. Nowadays, Deep Belief Networks (DBN) automatically is able to find out the representation of speech signal. This paper tries to approach a structure optimization of DBN which based on the combined technique of evolutionary computation to enhance the single speaker speech. It firstly extracts from the feature of speech signal then applies them to construct lots of random subspaces. The result of the conducted experimental in the evolutionary computation of DBN indicates the structure have an improvement for speech recognition.

show abstract

“…1. The outcomes of auditory comb filter hypothesis [12] and SPADE analysis in noisy speech recognition [11,14] have strongly inspired to develop this technique. The hypothesis can be implemented in WERB-SPADE by using comb filters and ERB like WP decomposition of acoustic speech signal.…”

Section: Sub-band Periodic and Aperiodic Decompositionmentioning

confidence: 99%

“…They have evaluated the performance of SPADE with AURORA-2J database in the presence of noise and claimed that proposed features have outperformed MFCC. Later Ishizuka and Nakatani [14] have expanded the SPADE analysis in frequency domain and have proposed new feature extraction technique named SPADE freQUEncy domain ENhancement (SPADE-QUEEN). They combined their proposed front end technique with different noise compensation techniques such as, spectral subtraction or Wiener filtering and studied the performance of robust front end technique with AURORA 2J database.…”

Section: Introductionmentioning

confidence: 99%

“…Here in this work we tried to enhance the performance of WERBC for Hindi voiced phoneme class. Motivated by the performance of WP based feature and SPADE analysis [9,11,14], here we discuss a WP based feature extraction technique named Wavelet ERB Subband based Periodicity and Aperiodicity Decomposition (WERB-SPADE). Wavelet packet (WP) decomposition is carried out similar to the Equivalent Rectangular Bandwidth (ERB) scale frequency resolution [18].…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Hindi phoneme classification using Wiener filtered wavelet packet decomposed periodic and aperiodic acoustic feature

Biswas

Sahu

Bhowmick

et al. 2015

Computers & Electrical Engineering

View full text Add to dashboard Cite

A feature extraction method using subband based periodicity and aperiodicity decomposition with noise robust frontend processing for automatic speech recognition

Cited by 17 publications

References 22 publications

Enhancement of speech perception in noise by periodicity processing: A neurobiological model and signal processing algorithm

Enhancement of speech perception in noise by periodicity processing: A neurobiological model and signal processing algorithm

Enhancing Single Speaker Recognition Using Deep Belief Network

Hindi phoneme classification using Wiener filtered wavelet packet decomposed periodic and aperiodic acoustic feature

Contact Info

Product

Resources

About