A psychoacoustic spectral subtraction method for noise suppression in automatic speech recognition

The basic problem suffered by a communication system is in the removal of additive background noise.Spectral Subtraction method (SSM) has been successfully implemented to suppress the background noise.This paper proposed a novel hybrid system for enhancing noise-corrupted speech which involves a parallel combination of Spectral Subtraction and soft thresholding method in wavelet domain for speech enhancement.The performance of hybrid system is compared with both Spectral Subtraction method and Wavelet thresholding technique (WTT) and found to have better SNR improvement.Implementation and evaluation have been made using MATLAB.

The spectral subtraction method is a conventional approach for single channel speech enhancement. The basic principle of this method is to estimate the short-time spectral magnitude of speech by subtracting estimated noise from the noisy speech spectrum and to combines it with the phase of the noisy speech. Besides reducing the noise, this method generates an unnatural and unpleasant noise, called remnant noise. This paper proposes a novel algorithm to reduce the remnant noise, and thus improving the overall quality of the enhanced speech. In this algorithm, the output of multi-band spectral subtraction (MBSS) method is used as the input signal again for next iteration process. After the MBSS method, the additive noise is changed to remnant noise. The remnant noise is re-estimated at each iteration. The new estimated noise, furthermore, is been used to process the next MBSS. This procedure is iterated a small number of times. The simulation results as well as informal subjective evaluations prove that the speech enhanced by proposed algorithm is more pleasant to listeners than the conventional MBSS algorithm. This reveals that the proposed algorithm reduces remnant noise satisfactorily and produces good speech quality with improved signal-to-noise ratio.

Rindik is a traditional music instrument originated from Bali that consists of 11 bamboo rods and played by a person hitting the bamboo rod with a rubber mallet in each player’s hands. Documentation of Rindik songs with automatic music transcription work is easier to do by separating it first. To overcome that challenge, two Rindik rods sound were separated using the spectral subtraction method. The noise spectrum is the spectrum of a single rod sound that needs to be muffled. The resulting audio is the other single rod sound and vice versa to get both single rod sound. The data consisted of single rod hit sound recordings of 11 single rod sound and 55 combinations of two-rods sounds being hit at the same time. The performance of the spectral subtraction method in separating Rindik sound was measured with MSE and SIR and also heard the noise exists in the separated audio signal. The experiment demonstrated a state-of-the-art performance consists of spectral subtraction with squared noise average magnitude with average MSE value 0.0126 and average SIR value 55.68 dB.

An acoustic noise suppression algorithm has been developed which suppresses noise from speech by first filtering it into a set of signals which approximate the loudness components perceived by the auditory system. These signals are generated by passing the input stimulus waveform through a filter bank with frequency bandwidths which approximate the ear's critical bandwidths. The noise on each signal is then suppressed using spectral subtraction techniques in a domain of simulated perception. This approach to noise suppression retains the intelligibility produced by spectral subtraction methods while eliminating the accompanying musical quality.

In this paper, we provide a new theoretical analysis of the amount of musical noise generated via generalized spectral subtraction based on higher order statistics. Power spectral subtraction is the most commonly used spectral subtraction method, and in our previous study a musical noise assessment theory limited to the power spectral domain was proposed. In this paper, we propose a generalization of our previous theory on spectral subtraction for arbitrary exponent parameters. We can thus compare the amount of musical noise between any exponent domains from the results of our analysis. We also clarify that less musical noise is generated when we choose a lower exponent spectral domain; this implies that there is no theoretical justification for using power/amplitude spectral subtraction.

This paper describes a method for enhancing speech corrupted by broadband noise. The method is based on the spectral noise subtraction method. The original method entails subtracting an estimate of the noise power spectrum from the speech power spectrum, setting negative differences to zero, recombining the new power spectrum with the original phase, and then reconstructing the time waveform. While this method reduces the broadband noise, it also usually introduces an annoying "musical noise". We have devised a method that eliminates this "musical noise" while further reducing the background noise. The method consists in subtracting an overestimate of the noise power spectrum, and preventing the resultant spectral components from going below a preset minimum level (spectral floor). The method can automatically adapt to a wide range of signal-to-noise ratios, as long as a reasonable estimate of the noise spectrum can be obtained. Extensive listening tests were performed to determine the quality and intelligibility of speech enhanced by our method. Listeners unanimously preferred the quality of the processed speech. Also, for an input signal-to-noise ratio of 5 dB, there was no loss of intelligibility associated with the enhancement technique.

To enable accurate exchange of information or data, the quality and intelligibility of speech play an important factor. However, pragmatically the quality of speech is distorted by the presence of background noise. This in turn leads to poor performance of the system as well as causes listener fatigue. To resolve such ambiguity we apply speech enhancement algorithms which improve the intelligibility as well as quality of speech which has been degraded by background noise. In this paper two such algorithms have been discussed for noise reduction. One is conventional spectral subtraction method and the other is the proposed modified version of spectral subtraction. While the former aims to improve speech quality degraded by additive background noise, the latter aims to improve speech quality from dynamic additive noise as well. Modified noise reduction algorithm is compared to conventional spectral subtraction based on SNR improvement introduced by them.

In this paper, some classical and modern spectral estimation algorithms are incorporated in the conventional spectral subtraction method for improving the performance of speech enhancement. There are three methods are proposed in this paper: the SSW method, the SSY method and the SSM method. Some typical environment noise signals are chosen from the common noise group. The noisy speech signals adopted in the experiments are the clean speech contaminated by the additive noise and the convolution noise at different signal-to-noise ratios respectively. The Itakura-Saito spectrum distance between the enhanced speech and the clean speech is served as the measurement of the quality of the speech enhancement system. Each type of the noisy speech signals will be enhanced by the conventional spectrum subtraction method and the three proposed methods respectively. Experimental results demonstrated the superior performance of the proposed methods over the conventional spectrum subtraction method. It is concluded that each of them is adept to manipulating different type of noises according to the corresponding characteristics.

An Iterative Graph Spectral Subtraction Method for Speech Enhancement

Combined speech enhancement and auditory modelling for robust distributed speech recognition

본 논문에서는 먼저 신경회로망의 학습에 오차역전파 학습 알고리즘을 사용하여 각 프레임에서의 음성 및 잡음 구간의 검출에 의한 음성인식 알고리즘을 제안한다. 그리고 신경회로망에 의하여 음성 및 잡음 구간의 검출에 따라서 각 프레임에서 잡음을 제거하는 스펙트럼 차감법을 제안한다. 본 실험에서는 제안한 음성인식알고리즘의 성능을 원음성에 백색잡음 및 자동차 잡음을 부가하여 인식율을 평가한다. 또한 인식시스템에 의하여 검출된 음성 및 잡음 구간을 이용하여 각 프레임에서의 스펙트럼 차감법에 의한 잡음제거의 실험결과를 나타낸다. 잡음에 의하여 오염된 음성에 대하여 신호대잡음비를 사용하여 본 알고리즘이 유효하다는 것을 확인한다. This paper first proposes the speech recognition algorithm by detection of the speech and noise sections at each frame using a neural network training by back-propagation algorithm, then proposes the spectral subtraction method which removes the noises at each frame according to detection of the speech and noise sections. In this experiment, the performance of the proposed recognition system was evaluated based on the recognition rate using various speeches that are degraded by white noise and car noise. Moreover, experimental results of the noise reduction by the spectral subtraction method demonstrate using the speech and noise sections detecting by the speech recognition algorithm at each frame. Based on measuring signal-to-noise ratio, experiments confirm that the proposed algorithm is effective for the speech by corrupted the noise using signal-to-noise ratio.

Vibration-based Structural Health Monitoring (SHM) systems use field measurements of operational signals, which are distorted by noise from many sources. Reducing this noise allows a more accurate assessment of the original "clean" signal and improves analysis results. The implementation of a noise reduction methodology for the Modal Distribution Method (MDM) is reported here. The spectral subtraction method is a popular broadband noise reduction technique used in speech signal processing. Its basic principle is to subtract the magnitude of the noise from the total noisy signal in the frequency domain. The underlying assumption of the method is that noise is additive and uncorrelated with the signal. In speech signal processing, noise can be measured when there is no signal. In the MDM, however, the magnitude of the noise profile can be estimated only from the magnitude of the Power Spectral Density (PSD) at higher frequencies than the frequency range of the true signal associated with structural vibrations under the additional assumption of white noise. The implementation of the spectral subtraction method to MDM may decrease the energy of the individual mode. In this work, a modification of the spectral subtraction method is introduced that enables the conservation of the energies of individual modes. The main difference is that any (negative) bars with a height below zero after subtraction are set to the absolute value of their height. Both noise reduction methods are implemented in the MDM, and an application example is presented that demonstrates its effectiveness when used with a signal corrupted by noise.

The Internet Protocol (IP) environment poses two relevant sources of distortion to the speech recognition problem: lossy speech coding and packet loss. In this paper, we propose a new front-end for speech recognition over IP networks. Specifically, we suggest extracting the recognition feature vectors directly from the encoded speech (i.e., the bit stream) instead of decoding it and subsequently extracting the feature vectors. This approach offers two significant benefits. First, the recognition system is only affected by the quantization distortion of the spectral envelope. Thus, we are avoiding the influence of other sources of distortion due to the encoding-decoding process. Second, when packet loss occurs, our front-end becomes more effective since it is not constrained to the error handling mechanism of the codec. We have considered the ITU G.723.1 standard codec, which is one of the most preponderant coding algorithms in voice over IP (VoIP) and compared the proposed front-end with the conventional approach in two automatic speech recognition (ASR) tasks, namely, speaker-independent isolated digit recognition and speaker-independent continuous speech recognition. In general, our approach outperforms the conventional procedure, for a variety of simulated packet loss rates. Furthermore, the improvement is higher as network conditions worsen.

This paper presents a comparative analysis of speech enhancement algorithms based on speech sparsity features. The sparsity of a speech is a signal-specific characteristic that plays a crucial role in identifying speech signal components from noisy speech. Conventional enhancement techniques like Spectral subtraction, Sub-space and statistical methods require knowledge of noise distribution for better performance. The paper compares major sparsitybased enhancement by highlighting the merits and demerits by comparing quality and intelligibility. An objective evaluation of quality and intelligibility has been performed to show relative performance on a common database. Results show that dictionary-based speech enhancement aided with voice activity detection and spectral subtraction performs better than others. The performances improve further if noise-type and noise dictionary is known as apriori in the joint dictionary and nonnegative matrix factorization method. This analysis wm help the researcher understand sparsity-based enhancement techniques and develop techniques to overcome existing issues.

A robust feature extraction technique using Teager Energy Operator (TEO) for Isolated Word Recognition (IWR) has been proposed in this paper. A feature extraction algorithm is motivated by the enhanced discrimination capability TEO that estimates the true energy of the source of a resonance. The robustness is further added using Cepstral Mean Normalization (CMN) on the estimated features. The robust features are computed from the speech signal of a given frame through a series of steps. First, the short time spectrum of each frame of speech signal is calculated. Second, the frame spectrum is passed through a Mel scaled triangular filter bank. Then, the average of absolute values of sequence obtained after applying TEO on each filter output is estimated. Finally, the cepstral coefficients are extracted by applying discrete cosine transform on the estimated averages. These coefficients are further normalized using CMN to get the final features denoted as Normalized Teager Energy Coefficient (NTEC) features. The effectiveness of this technique has been tested on TI-20 isolated word database in presence of white noise. The experimental results show the superiority of the proposed technique over conventional MFCC, Spectral Subtraction (SS) and CMN methods.