Filtered Speech Signal Research Articles

Articulation Battery Test For Tamil Speech Disorder Signals

Articulation disorder is referred as difficulty occurs in the pronunciation of specific speech sounds. An irregular coordination of the movement of tongue, lips, palate, jaw, respiratory system, vocal tract, height of the larynx, air flow through nasal leads to the incorrect production of speech sounds. The objective of this paper is to propose a computational model based on Recurrent Neural Network (RNN) algorithm to categorize the phonological patterns of Tamil speech articulation disorder signals into four predefined groups, namely, substitution, omission, distortion and addition. The methodology of the proposed work is described as follows. (1) List of articulation disorder test words suggested by Speech Language Pathologists (SLPs) is selected for this experimental study. (2) Real time speech signals that comprise of Tamil vowels (Uyir eluthukkal) and consonants (Meiyeluthukkal) are collected from people with articulation disorder. (3) Acoustic noise and weak signals are eliminated by applying Low pass filter to acquire the filtered speech signal. (4) Mel-Frequency Cepstral Coefficients (MFCCs) technique is implemented to extract the prominent features from denoised signals. (5) Principal Component Analysis (PCA) method is employed to choose fine-tune feature subset. (6) The refined features are employed to calibrate RNN model for classification. Results show that RNN model achieves 90.25% classification accuracy when compared to other artificial neural network algorithms.

Monitoring Cognitive Workload Using Vocal Tract and Voice Source Features

Monitoring cognitive workload from speech signals has received much attention from researchers in the past few years as it has the potential to improve performance and fidelity in human decision making. The bulk of the research has focused on classifying speech from talkers participating in cognitive workload experiments using simple reading tasks, memory span tests and the Stroop test, typically into three levels of low, medium and high cognitive workload. This study focuses on using parameters extracted from the vocal tract and the voice source components of the speech signal for cognitive workload monitoring. The experiment used in this study contains 98 participants, the levels were obtained by using a reading task and three Stroop tasks which were randomly ordered for each participant and an adequate rest time was used inbetween tasks to mitigate the effect of cognitive workload from one task affecting the subsequent one. Vocal tract features were obtained from the first three formants and voice source features were extracted using signal analysis on the inverse filtered speech signal. The results show that on their own, the vocal tract features outperform the voice source features. The MCR of 33.92% ± 1.05 was achieved with a SVM classifier. A weighted combination of vocal tract and voice source features classified with SWM classifier fused at the output level achieved the lowest MCR of 32.5%.

Voice/non-voice detection using phase of zero frequency filtered speech signal

Voice/non-voice detection refers to the task of detecting the presence or absence of vocal folds activity regions in the speech signal. Most of the existing state-of-the-art methods depend exclusively on the amplitude of the signal either in time or frequency domains, and their performance is significantly affected for weakly voiced, laryngeal transitions and noisy segments of speech. In this paper, we propose a robust method for detecting voice/non-voice regions in the speech signal based on the harmonics of the phase of the source signal. Here, the source signal is derived by removing the effect of vocal tract resonances from the speech signal by using zero frequency filtering (ZFF). The experimental results demonstrate the robustness of the proposed method for accurate detection of voiced/non-voiced regions in the speech signal during adverse conditions. The performance of the proposed method is compared with one of the state-of-the-art methods based on sum of residual harmonics, and three well known standard voice activity detection (VAD) algorithms: G729B, adaptive multi-rate VAD option-1 (AMR1) and adaptive multi-rate VAD option-2 (AMR2).

The effects of frequency lowering on speech perception in noise with adult hearing-aid users

Objective: Frequency lowering (FL) strategies move high frequency sound into a lower frequency range. This study determined if speech perception differences are observed between some of the different frequency lowering strategies that are available. Design: A cross-sectional, repeated-measures design was used to compare three hearing aids that used wide-dynamic range compression (WDRC) and either non-linear frequency compression (NFC), linear frequency transposition (LFT), or frequency translation (FT). The hearing aids were matched to prescriptive real ear targets for WDRC. The settings for each FL strategy were adjusted to provide audibility for a 6300 Hz filtered speech signal. Sentence recognition in noise, subjective measures of sound quality, and a modified version of the speech intelligibility index (SII) were measured. Study sample: Ten adults between the ages of 63 to 82 years with bilateral, high frequency hearing loss. Results: LFT and FT led to poorer sentence recognition compared to WDRC for most individuals. No difference in sentence recognition occurred with and without NFC. The quality questionnaire and SII showed few differences between conditions. Conclusion: Under similar fitting and testing conditions of this study, FL techniques may not provide speech understanding benefit in certain background noise situations.

Efficient environmental noise reduction system for speech enhancement via fuzzy-based extreme learning machine and fuzzy adaptive Wiener filter

This paper proposes the application of fuzzy-based extreme learning machine (FELM) and fuzzy-based adaptive Wiener filter for background noise reduction system to improve the signal to noise ratio (SNR) and to reduce the minimum mean square error (MMSE). The Wiener filter minimizes the mean square error and it provides better performance than the conventional filters. Though the background noise is uncertain, fuzzy inference systems (FIS) are proposed in ELM to classify the background noises and Wiener filters are proposed to update Wiener filter coefficients that will increase the SNR of the filtered speech signal. Fuzzy adaptive Wiener filter depends on the adaptation of the filter transfer function from sample to sample based on the signal statistics and it is implemented in time domain rather than in frequency domain to accommodate for the varying nature of the speech signal. The proposed FELM is compared with ELM and fuzzy radial basis function network (FRBFN) for noise classification and fuzzy adaptive Wiener filter is compared with Wiener and adaptive Wiener filters for noise cancellation. Simulation results show that FELM improves the percentage of classification by 9% than FRBFN, 4% than ELM and fuzzy adaptive Wiener filter improves the SNR by 8 dB than the conventional Wiener filter. The real-time implementation of the system is done using TMS320C6713 DSK starter kit. The real-time practical setup using DSK shows an improved SNR of 5 dB.

Environmental noise reduction system using fuzzy neural network and adaptive fuzzy algorithms

This article proposes the application of fuzzy based radial basis function network (FRBFN) and fuzzy based adaptive Wiener filter for background noise reduction system to improve the signal-to-noise ratio (SNR) and to reduce the minimum mean square error (MMSE). The Wiener filter works as to minimise the mean square error, and it provides better performance than the conventional filters. Though the background noise is uncertain, fuzzy inference systems are proposed in RBFN to classify the background noises and in Wiener filter to update Wiener filter coefficients that will increase the SNR of the filtered speech signal. The proposed FRBFN is compared with RBFN for noise classification, and fuzzy adaptive Wiener filter is compared with Wiener and adaptive Wiener filters for noise cancellation. Simulation result shows that FRBFN improves the percentage of classification by 7% than RBFN, and fuzzy adaptive Wiener filter improves the SNR by 6 dB than the conventional Wiener filter. The real time implementation of the system is done using TMS320C6713 DSK starter kit. The real time practical setup using DSK shows an improved SNR of 4 dB.

응급상황에서의 음성인식을 위한 필터기 구현

일반적으로 음성인식 시스템의 사용에 가장 저해되는 요소에는 배경 잡음을 들 수 있다. 잡음은 음성인식 시스템의 성능을 저하시키고, 이로 인해 사용 장소의 제약을 많이 받게 되는 이유가 된다. 이런 잡음의 영향을 해결하기 위해 본 논문에서는 음질 향상에 목적을 두고 신호단계에서부터 잡음성분을 제거하는 필터 중 FIR필터의 대역통과를 이용하여 일반적으로 사람의 음성 주파수 영역과 잡음 영역을 추출한 정보를 토대로 Wiener 필터를 구현, 그 성능을 향상하여, 전송되어지는 음성신호구간에서 잡음구간과 음성구간에 따라 잡음을 유연하게 처리하도록 구현하였다. Generally, the mal factor for speech recognition is the background noise in speech recognition. The noise is the reason to reduce the speech recognition performance. Owing to the fact, the place to recognize is very important. To improve the recognition performance from the sound having noise, we implemented the noise filtered Wiener filter at the signal process step which adopted the FIR filter. In FIR filter, it deal with the filtered speech signal which is appropriate frequency range of human speech frequency range. Therefore, we make the recognition system distinguish between noise and speech sound from the incoming speech signal.

Voiced/Nonvoiced Detection Based on Robustness of Voiced Epochs

In this paper, a new method for voiced/nonvoiced detection based on epoch extraction is proposed. Zero-frequency filtered speech signal is used to extract the instants of significant excitation (or epochs). The robustness of the method to extract epochs in the voiced regions, even with small amount of additive white noise, is used to distinguish voiced epochs from random instants detected in nonvoiced regions. The main feature of the proposed method is that it uses the strength of glottal activity as against using the periodicity of the signal. Performance of the proposed algorithm is studied on TIMIT and CMU ARCTIC databases, for two different noise types, white and vehicle noise from the NOISEX database, at different signal-to-noise ratios (SNRs). The proposed method performs similar or better than the popular normalized crosscorrelation based voiced/nonvoiced detection used in the open source utility <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">wavesurfer</i> , especially at lower SNRs.

Method of estimating pitch by using ratio of maximum peak to candidate for maximum of autocorrelation function and device using the method

A device and a method for estimating an open-loop pitch in a general speech CODEC are disclosed. The open-loop pitch estimation device includes an autocorrelation function calculation unit which calculates a normalized autocorrelation function from a perceptual weighing filtered speech signal, a maximum autocorrelation function and lag estimation unit which estimates a maximum autocorrelation function and candidates for the maximum autocorrelation function, a pitch candidate decision unit which decides candidates for a pitch by using the ratio of the estimated maximum autocorrelation function to the candidates for the estimated maximum autocorrelation function, and lags of which values are smaller than a predetermined threshold value, and a pitch estimation unit which estimates a pitch between the candidates for a pitch and the lags corresponding to the estimated maximum autocorrelation function by using a pitch of a previous frame of the speech signal.

Noise reduction in subbanded speech signals

The presence of speech in a filtered speech signal is detected for the purpose of suspending noise level calculations during periods of speech. A received speech signal is split into a plurality of subband signals. A subband variable gain is determined for each subband based on an estimation of the noise level in the received voice signal and on an envelope of the received signal in each subband. Each subband signal is multiplied by the subband variable gain for that subband. The subband signals are combined to produce an output voice signal.

A wavelet transform approach to blind adaptive filtering of speech from unknown noises

This paper is concerned with the application of wavelet transforms (WT) to blind adaptive filtering (BAF) of speech signals from unknown colored noise under a single-microphone constraint when neither speech nor noise are separately accessible. The paper describes the BAF system which performs discrete WT of noisy speech signals and identifies and separates noise from speech. The BAF then tunes sets of weights that multiply WT coefficients on different scales of the WT to produce a significantly improved signal-to-noise ratio and intelligibility of the filtered speech signal at the output of the BAF.

On decomposing speech into modulated components

We model a segment of filtered speech signal as a product of elementary signals as opposed to a sum of sinusoidal signals. Using this model, one can better appreciate the basic relationships between envelopes and phases or instantaneous frequencies (IFs) of signals. These relationships reveal some interesting properties of the signal's modulations. For instance, if the contribution due to a signal's envelope, specifically the Hilbert transform of its log-envelope, is removed from the signal's phase, then the resulting signal's IF is strictly positive. In addition, filtered speech signal having a bandwidth of B Hz can be essentially represented by the log-envelope and IF that have the same B Hz bandwidths. We extend the above ideas to decompose speech into modulated components. Specifically, a bank of data-adaptive filters (in a cross-coupled configuration) are used to decompose speech into its components; each adaptive filter is a simple single resonance bandpass filter (whose center-frequency or pole-location closely follows the desired formant frequency) supplemented by an adaptive all-zero filter (whose zero-locations sufficiently reduce unwanted leakage from neighboring formants). The filtered components are then represented by their respective log-envelopes and positive IFs; these small number of modulations closely approximate the speech signal.

Method and system for adaptive filtering of speech signals using signal-to-noise ratio to choose subband filter bank

A method and system for adaptively filtering a speech signal. The method includes decomposing the signal into subbands, which may include performing a discrete Fourier transform on the signal to provide approximately orthogonal components. The method also includes determining a speech quality indicator for each subband, which may include estimating a signal-to-noise ratio for each subband. The method also includes selecting a filter for filtering each subband depending on the speech quality indicator, which may include estimating parameters for the filter based on a clean speech signal. The method further includes determining an overall average error for the filtered subbands, which may include calculating a mean-squared error. The method still further includes identifying at least one filtered subband which, if excluded from the filtered speech signal, would reduce the overall average error determined, and combining, with exception of the filtered subbands identified, the filtered subbands to provide an estimated filtered speech signal. The system includes filters and software for performing the method.

Chapter 26 A digital speech processor and various speech encoding strategies for cochlear implants

We distinguish two main categories in the speech decoding strategies: those based on feature extraction and those based on time information of the filtered incoming signal. In the first group only those electrode pairs corresponding to the localisation of the four or more, maximum spectral peaks or the formants of the speech signal are stimulated. The second category is essentially based on the timing information that is included in the filter outputs of the filtered speech signal. The Laura implant has the capability to be programmed to the different decodings-strategies. To understand the potential possibilities the main features of the implant are discussed.

Single Band Amplitude Envelope Cues as an Aid to Speechreading

Amplitude envelopes derived from speech have been shown to facilitate speech-reading to varying degrees, depending on how the envelope signals were extracted and presented and on the amount of training given to the subjects. In this study, three parameters related to envelope extraction and presentation were examined using both easy and difficult sentence materials: (1) the bandwidth and centre frequency of the filtered speech signal used to obtain the envelope; (2) the bandwidth of the envelope signal determined by the lowpass filter cutoff frequency used to "smooth" the envelope fluctuations; and (3) the carrier signal used to convey the envelope cues. Results for normal hearing subjects following a brief visual and auditory-visual familiarization/training period showed that (1) the envelope derived from wideband speech does not provide the greatest benefit to speechreading when compared to envelopes derived from selected octave bands of speech; (2) as the bandwidth centred around the carrier frequency increased from 12.5 to 1600 Hz, auditory-visual (AV) performance obtained with difficult sentence materials improved, especially for envelopes derived from high-frequency speech energy; (3) envelope bandwidths below 25 Hz resulted in AV scores that were sometimes equal to or worse than speechreading alone; (4) for each filtering condition tested, there was at least one bandwidth and carrier condition that produced AV scores that were significantly greater than speechreading alone; (5) low-frequency carriers were better than high-frequency or wideband carriers for envelopes derived from an octave band of speech centred at 500 Hz; and (6) low-frequency carriers were worse than high-frequency or wideband carriers for envelopes derived from an octave band centred at 3150 Hz. These results suggest that amplitude envelope cues can provide a substantial benefit to speechreading for both easy and difficult sentence materials, but that frequency transposition of these signals to regions remote from their "natural" spectral locations may result in reduced performance.

Amplitude envelopes and speechreading

Amplitude envelopes derived from filtered bands of speech provide varying degrees of benefit to speechreading. Three parameters related to envelope extraction are examined using both easy and difficult sentence materials: (1) the bandwidth and center frequency of the filtered speech signal used to obtain the envelope, (2) the bandwidth of the envelope signal, and (3) the carrier signal used to convey the envelope. Thus far, results for normal hearing subjects presented with difficult speech materials suggest that: (1) the envelope of wideband speech did not always provide the greatest benefit to speechreading when compared to envelopes derived from octave bands of speech; (2) as the bandwidth of the envelope signal increased from 100–4000 Hz, auditory‐visual (AV) performance improved slightly, whereas bandwidths below 100 Hz resulted in AV scores that were equal to or worse than speechreading alone; and (3) low‐frequency carrier signals were better than high‐frequency or wideband carrier signals, at least for envelopes derived from an octave band of speech centered at 500 Hz. Similar tests with easy speech materials are in progress. [Work supported by NIH.]

The deterioration of hearing with age: Frequency selectivity, the critical ratio, the audiogram, and speech threshold

The frequency selectivity of the auditory system was measured by masking a sinusoidal signal (0.5, 2.0, or 4.0 kHz) or a filtered-speech signal with a wideband noise having a notch, or stopband, centered on the signal. As the notch was widened performance improved for both types of signal but the rate of improvement decreased as the age of the 16 listeners increased from 23 to 75 years, indicating a loss in frequency selectivity with age. Auditory filter shapes derived from the tone-in-noise data show (a) that the passband of the filter broadens progressively with age, and (b) that the dynamic range of the filter ages like the audiogram. That is, the range changes little with age before 55, but beyond this point there is an accelerating rate of loss. The speech experiment shows comparable but smaller effects. The filter-width measurements show that the critical ratio is a poor estimator of frequency selectivity because it confounds the tuning of the system with the efficiency of the signal-detection and speech-processing mechanisms that follow the filter. An alternative, one-point measure of frequency selectivity, which is both sensitive and reliable, is developed via the filter-shape model of masking.

Perception of slope-overload distortion in delta-modulated speech signals

An experiment was performed in which controlled amounts of slope-overload distortion were produced in specified sections of a test sentence. Preliminary trials showed that the commonly used measure "percent of time of slope overload" is not a good indicator of perceptual distortion in delta-modulated speech signals. A more appropriate parameter for specifying slope-overload distortion is the rms slope truncation, which is the rms value of the truncated portion of the time derivative of the signal beyond the maximum absolute slope of the delta modulator. For a 3 kHz low-pass filtered speech signal, the just-perceptible level of distortion for the most sensitive of three subjects was 1.8 quantization units/20 kHz sampling period where 1 quantization unit = 2 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-11</sup> of the peak-to-peak range of the waveform. This figure varied slightly for different speech sounds, but the range of variation was substantially less than the differences between subjects.