Unvoiced Signals Research Articles

Signal Processing: New Stochastic Feature of Unvoiced Pronunciation for Whisper Speech Modeling and Synthesis

Whisper is an indispensable way in speech communication, especially for private conversation or human-machine interaction in public places such as library and hospital. Whisper is unvoiced pronunciation, and voiceless sound is usually considered as noise-like signals. However, unvoiced sound has unique acoustic features and can carry enough information for effective communication. Although it is a significant form of communication, currently there is much less research work on whisper signal than common speech and voiced pronunciation. Our work extends the research of unvoiced pronunciation signal by introducing a novel signal feature, which is further applied in unvoiced signal modeling and whisper sound synthesis. The statistics of amplitude for each frequency component is studied individually, based on which a new feature of “consistent standard deviation coefficient” is revealed for the amplitude spectrum of unvoiced pronunciation. A synthesis method for unvoiced pronunciation is proposed based on the new feature, which is implemented by STFT with artificially generated short-time spectrum with random amplitude and phase. The synthesis results have identical quality of auditory perception as the original pronunciation, and have similar autocorrelation as that of the original signal, which proves the effectiveness of the proposed stochastic model of short-time spectrum for unvoiced pronunciation

Implementation of ZCR and STE techniques for the detection of the voiced and unvoiced signals in Continuous Punjabi Speech

Implementation of ZCR and STE techniques for the detection of the voiced and unvoiced signals in Continuous Punjabi Speech

Voice Activity Detection Using an Improved Unvoiced Feature Normalization Process in Noisy Environments

Noise-elimination technology is used to eliminate noise, including environmental noise, from voice signals in order to increase voice recognition rates. Noise estimation is the most important factor in noise-elimination technology. One of the effective estimation methods is voice activity detection, which is based on the statistical properties of noise and voice. This method is a way of estimating noise using the statistical properties of both noise and voice, which have an independent Gaussian distribution. In cases of severe differences in a statistical property, like white noise, the method is very reliable but limited to signals having a low signal-to-noise ratio (SNR) or having speech shape noise, which has statistical properties similar to voice signals. Methods to increase the voice recognition rate suffer from decreasing voice recognition performance due to distortion of the voice spectrum and to missing voice frames, because noise remains if there has been incorrect estimation of the noise. Degradation in voice recognition performance emerges in the differences between the model training environment and the voice recognition environment. In order to decrease environmental discordance, various silence feature normalization methods are used. Existing silence feature normalization suffers from degradation of recognition performance because the classification accuracy for the voiced and unvoiced signals decreases by an increasing energy level in the silence section of a low SNR. This paper proposes a robust voice characteristic detection method for noisy environments using feature extraction and unvoiced feature normalization for a classification relative to the voiced and unvoiced signals. The suggested method constitutes a model for recognition by extracting the characteristics for classification of the voiced and unvoiced signals in a high SNR environment. Also, the model affects noise for voice characteristics less, and recognition performance improves by using the Cepstrum feature distribution property of voiced and unvoiced signals with a low SNR. The model was checked for its ability to improve recognition performance relative to the existing method based on recognition experiment results.

A coder with adaptive filters and non-uniform techniques for simple speech communication

This paper proposes a new method of non-uniform sampling coding that utilizes a modified zero-crossing rate filter and a variable filter. In speech signals, the modified zero-crossing rate refers to the number of times in which the signal crosses zero in a single frame; the rate is high in noisy sections but low in sections of vocal sounds that are quasi-periodic. And the modified zero-crossing rates of unvoiced signals are similar to those of voiced signals. In speech signals, the non-uniform sampling coding method utilizes the peaks and valleys in a frame. In noisy sections, there are a lot of zero crossings and thus creating many peaks and valleys, while sections with voice sounds have zero-crossings numbering only 2–3 times of their fundamental frequency. The method proposed in this paper utilizes the number of peaks and valleys in determining the validity of a sample; the less we have peaks and valleys, the higher compression rates we get. Therefore, by decreasing the number of samples from sections with noise and voiceless sounds which have a high number of peaks and valleys, and by creating a signal that has little effect on the recognized signals, we may acquire a better compression rate. As examined and proposed in this paper, these features can be used to design a filter coefficient for a better performance in order to improve compression rates of non-uniform sampling coding and to maintain high quality signals.

PROPOSED A NEW APPROACH FOR VOICED / UNVOICED DECISION OF SPEECH FILE USING LAGRANGE TECHNIQUE

In speech analysis, voiced/unvoiced decision is usually performed in extracting the information from the speech signals. This paper presented a new approach for voiced/unvoiced (V/UV) decision; the approach is used Lagrange polynomial interpolation for voiced/ unvoiced signals speech decision, depending on the advantage of the characteristics of the interpolation techniques. Speech file is divided in different size of blocks, each block is evaluated by Lagrange technique, and the results are evaluated by one of most popular voiced/unvoiced measured which is the short time energy classifier. The results show that Lagrange technique worked well as classifier and give an accepted decision to analysis the speech to voice and unvoiced sample, especially with small size non overlapping blocks.

Speech enhancement using hybrid gain factor in critical-band-wavelet-packet transform

Noise masking threshold (NMT) has been extensively applied in wavelet-based speech enhancement. A gain factor is typically derived according to the NMT to suppress the additive noise. This investigation proposes a gain factor in each wavelet subband subject to a perceptual constraint. This perceptual constraint preserves the wavelet coefficients (WCs) of noisy speech when the level of residual noise is smaller than the NMT. If the residual noise level exceeds the NMT, then the wavelet coefficients of noisy speech are suppressed to reduce the corrupting noise. A lower bound on the gain factor is also proposed to prevent low-SNR regions, such as unvoiced signal, from being over-attenuated. Experimental results show that the proposed approach improves the naturalness, and does not cause annoying residual noise in the enhanced speech.

Speaker Identification Based on Incremental Learning Neural Network

Speech signal has various features of speakers. This feature is extracted from speech signal processing. The speaker is identified by the speaker identification system. In this paper, we propose the speaker identification system that uses the incremental learning based on neural network. Recorded speech signal through the microphone is blocked to the frame of 1024 speech samples. Energy is divided speech signal to voiced signal and unvoiced signal. The extracted 12 orders LPC cpestrum coefficients are used with input data for neural network. The speakers are identified with the speaker identification system using the neural network. The neural network has the structure of MLP which consists of 12 input nodes, 8 hidden nodes, and 4 output nodes. The number of output node means the identified speakers. The first output node is excited to the first speaker. Incremental learning begins when the new speaker is identified. Incremental learning is the learning algorithm that already learned weights are remembered and only the new weights that are created as adding new speaker are trained. It is learning algorithm that overcomes the fault of neural network. The neural network repeats the learning when the new speaker is entered to it. The architecture of neural network is extended with the number of speakers. Therefore, this system can learn without the restricted number of speakers.

Incremental Learning을 이용한 화자 인식

음성 속에는 화자의 특징이 포함되어 있다. 본 논문에서는 신경망에 기초한 Incremental Learning을 이용하여 화자 수에 제한 받지 않는 화자 인식 시스템을 제안한다. 컴퓨터를 통하여 녹음된 음성 신호는 End Detection과정을 통하여 유성음과 무성음을 분류하고 LPC를 이용해 12차수의 Cepstral Coefficients를 추출한다. 이 계수는 화자 식별을 위한 학습 입력값으로 사용 된다. Incremental Learning은 이미 학습한 Weight들을 기억하고 새로운 data에 대해서만 학습을 하는 학습 방법으로 Neural Network 구조가 화자 수에 따라 늘어나므로 화자 수에 제한을 받지 않고 학습이 가능하다. Speech signal has the features of speakers. In this paper, we propose the speaker identification system which use the incremental learning based on neural network. Recorded speech signal through the Mic is passed the end detection and is divided voiced signal and unvoiced signal. The extracted 12 order cpestrum are used the input data for neural network. Incremental learning is the learning algorithm that the learned weights are remembered and only the new weights, that is created as adding new speaker, are trained. The architecture of neural network is extended with the number of speakers. So, this system can learn without the restricted number of speakers.

Frequency selective harmonic coding

The present invention relates to a method of encoding speech comprised of processing the speech by harmonic coding to provide, a fundamental frequency signal, and a set of optimal harmonic amplitudes, processing the harmonic amplitudes, and the fundamental frequency signal to select a reduced number of bands, and to provide for the reduced number of bands a voiced and unvoiced decision signal, an optimal subset of magnitudes and a signal indicating the positions of the reduced number of bands, whereby the speech signal may be encoded and transmitted as the pitch signal and the signals provided for the reduced number of bands with a bandwidth that is a fraction of the bandwidth of the speech.

Encoding speech using prototype waveforms

Voiced speech is interpreted as a concentration of slowly evolving pitch-cycle waveforms. This signal can be reconstructed by interpolation from a downsampled sequence of pitch-cycle waveforms with a rate of one prototype waveform per 20-30 ms interval. The prototype waveform is described by a set of linear-prediction (LP) filter coefficients describing the formant structure and a prototype excitation waveform, quantized with analysis-by-synthesis procedures. The speech signal is reconstructed by filtering an excitation signal consisting of the concatenation of (infinitesimal) sections of the instantaneous excitation waveforms. To obtain the correct level of periodicity, the short-term and the long-term correlations between the instantaneous excitation waveforms can be controlled explicitly. Thus, distortions such as noise, reverberation, and buzziness can be prevented. The coding method is easily combined with existing LP-based speech coders, such as CELP, for unvoiced signals. Excellent voiced speech quality is obtained at rates between 3.0 and 4.0 kb/s. >

Finite state CELP for variable rate speech coding

The performance of a variable rate code excited linear predictor system is investigated. The coding system is based on a finite state CELP (FSCELP) frame work. Each individual state is primarily identified with a LPC model order, LPC coefficients bit allocation, excitation code book population density and state encoding rate. Successive input speech vectors are encoded at a rate that depends on the current state of the FSCELP system and the input vector characteristics. The use of a finite state system involves implicit clustering of speech signals. The lower rate states are selected during highly correlated steady state speech segments when relatively few bits are required to obtain adequate fidelity. For speech signals with a strong glottal excitation, unvoiced signals and transient speech segments, a relatively greater quantisation accuracy is needed to obtain good fidelity and therefore higher rate states of the system are used. Further improvement is obtained by using gamma populated excitation codebooks, for those states that are mainly used to encode speech signals with a strong underlying glottal excitation pulses. Experiments focus on investigation of the varying encoding requirements of the excitation signal for low pass, voiced, unvoiced and transient speech signals. The parameters of the finite state CELP system are designed to match the encoding requirements of typical speech signals. The greater part of the coding gain is obtained from variable rate encoding of the excitation signal. Using a six-state FSCELP, good quality speech is obtained at an average, maximum and minimum bit rates of 4 kbit/s, 10 kbit/s and 2 kbit/s, respectively.