This paper addresses the problem of harmonic distortion that introduced in traditional short time noise suppression methods, especially in low signal-to-noise ratio (SNR) noisy environments. We propose a simple but effective harmonic regeneration method to solve this problem. The full harmonics of the speech are obtained by applying a non-linear function to the enhanced speech after traditional short time noise suppression methods. The artificial signal is then refined by updating the spectral amplitude in three different frequency sub-bands respectively. The algorithm is theoretically analyzed and tried. Both the subjective and objective experiment results show that the harmonic regeneration algorithm can help improve the quality of the enhanced speech.

Speech enhancement using harmonic regeneration

A frame is a fmaily {f_i}_(i=1)~∞ of elements in a Hilbert space with the property that every element in can be written as a (infinite) linear combination of the frame elements. Frame theory describes how one can choose the corresponding coefficients, which are called frame coef- ficients. From the mathematical point of view this is gratifying, but for applications it is a problem that the calculation requires inversion of an operator on The projection method is introduced in order to avoid this problem. The basic idea is to con- sider finite subfamilies {f_i}_(i=1)~n of the frame and the orthogonal projection P_n onto its span. For f∈P_n f has a representation as a linear combination of f_i,i=1,2,…,n and the correspond- ing coefficients, can be calculated using finite dimensional methods. We find conditions implying that those coefficients converge to the correct frame coefficients as n→∞, in which case we have avoided the inversion problem. In the same spirit we approximate the solution to a moment prob- lem. It turns out, that the class of "well-behaving frames" are identical for the two problems we consider.

Riesz frames and approximation of the frame coefficients

In this paper, we introduce a practical real-time speech enhancement system based on the TI TMS320VC5509A digital signal processor (DSP). A stereo audio codec is used as the analog input/output (I/O). The algorithm for noise reduction is an improved and optimized spectral-subtraction method. The detailed design of the hardware and software is described and the experiments show an improved performance. Low power consumption and high efficiency make the whole system suitable for portable devices.

Low Power Embedded Speech Enhancement System Based on a Fixed-Point DSP

A new characterization of the generalized inverse using projections on level sets

Implementation of novel algorithms on Digital Signal Processing [DSP] processor to extract the speech signal from a noisy signal is always of immense interest. Speech signals are usually complex that requires processing of signal in short frames, and thus DSP processors are widely used to process the speech signals in mobile phones. The performance of these devices is comparatively very well in noisy conditions, as compared with traditional processors. The speech signal is degraded by either echo or background noise and as a result, there exists a requirement of digital voice processor for human–machine interfaces. The chief objective of speech enhancement algorithms is to improve the performance of voice communication devices by boosting the speech quality and increasing the intelligibility of voice signal. Popular speech enhancement algorithms rely on frequency domain approaches to estimate the spectral density of noise. This paper proposes a method of assessment, wherein frequency components of noisy speech signal are filtered out using multiband filters. The multiband filters are developed in C6748 DSP processor. Experimental results demonstrate an improved Signal to Noise Ratio [SNR] with fewer computations.

Implementation of Spectral Subtraction Using Sub-band Filtering in DSP C6748 Processor for Enhancing Speech Signal

As single channel speech enhancement, common me- thods based on short-time spectral amplitude estimation such as spectral-subtraction and wiener filtering are considered to be simple and effective for suppressing the additive noise in noisy environments. However, these techniques may distort the high frequency harmonics of speech, which are important to the lan- guage understanding. To solve this problem, we propose a new approach. It works by using a comb filter in time domain after wiener filtering. With the harmonics of voiced speech enhanced and most of the noise suppressed, another comb filter in frequen- cy domain is used to recalculate the suppression gain function, renewing the high frequency harmonics that have been com- pressed. This new method is theoretically derived, and experi- mentally demonstrates an improvement for harmonic enhance- ment and noise reduction. In this paper, we propose a new method. It uses a classical wiener filter first. Then a comb filter in time domain based on accurate pitch estimation is used to enhance the voiced speech harmonics and reduce the noise. After that, an improved har- monic regeneration method is processed in frequency domain to recover the harmonics which have been destroyed. The wiener filtering ensures the accuracy of pitch estimation for the comb filter, and the comb filtering further suppresses the resi- dual noise after wiener filtering and improves the performance of the harmonic recovering. This new method can be potential- ly combined with previous algorithms but performs better than either one separately for enhancing the sound quality and speech clarity. We believe the proposed method has both theo- retical and practical merits.

Harmonic Enhancement with Noise Reduction of Speech Signal by Comb Filtering

In this paper we show that frame coefficients related to Riesz bases of integer-translates are solutions of bi-infinite invertible Toeplitz systems. Hence the corresponding frame coefficients may be obtained by solving Toeplitz system directly, without complicated calculation of the inverse frame operator. We apply the finite section method to Toeplitz operator and provide a bi-directional Levinson algorithm to approximate solutions of biinfinite Toeplitz system, whose convergence is guaranteed by the recursive nature and properties of finite section method. Other advantages of the algorithm are the modest computational cost and less storage requirement.

A recursive algorithm approximating frame coefficients related to Riesz bases of translates

In this paper we present a new methodology for transversal filter synthesis based on frames of integer-translates. The method based on frame theory is an alternative approach to traditional transversal filter synthesis and very suitable for the tapped-delay structure. Furthermore we provide necessary and sufficient conditions for {f(·�k)}k∈Z to form a Riesz basis or frame sequence for l 2 (Z). If the translated versions of input signal form a Riesz basis, the bi-directional Levinson algorithm can be used to approximate filter coefficients directly, whose computational cost is modest and storage requirement is less. Besides, we extend finite transversal filter to infinite transversal filter which can be formulated exactly by systems of integer- translates and regard the former as a projection of the latter. Finally, we present the geometrical nature of finite transversal filter synthesis based on Riesz bases of translates.

Chaohuan Hou

Papers

Low Power Embedded Speech Enhancement System Based on a Fixed-Point DSP

Harmonic Enhancement with Noise Reduction of Speech Signal by Comb Filtering

A recursive algorithm approximating frame coefficients related to Riesz bases of translates

Transversal Filter Synthesis Based on Frames of Translates