Mohamed Krini

Spectral Refinement and its Application to Fundamental Frequency Estimation

In this paper a method for spectral refinement (SR) of speech and audio signals and its application to fundamental frequency estimation is presented. The SR procedure is applied as a post-processor on the output of a standard short-term frequency analysis. The algorithm is based on a linear combination of weighted subband signal vectors and thus has low computational complexity. The new scheme can be applied either as a refinement of only a subset of the frequency bands or as a refinement of the entire frequency range including the computation of additional frequency supporting points. Several algorithmic parts, e.g., noise suppression or fundamental frequency estimation, can achieve better results if a better resolution - at least in the lower frequency range - can be provided. In this contribution an enhanced fundamental frequency estimation method is proposed, that allows reliable operation at low signal-to-noise scenarios even for very low fundamental frequencies. Evaluations have shown that a significant improvement can be accomplished when utilizing the SR method as a pre-processor for fundamental frequency estimation.

Low-complexity noise power spectral density estimation for harsh automobile environments

In this paper a simple yet robust noise power spectral density estimation algorithm is presented. The motivation for the algorithm developed here is the harsh noise environment present in automobiles along with the need to keep the complexity low for real-time implementations. The scheme is based on a multiplicative estimator in which multiple increment and decrement time-constants are utilized. The time-constants are chosen based on noise-only and speech-like situations. Further, by observing the long-term “trend” of the noisy input spectrum, suitable time-constants are chosen which reduces the tracking delay significantly. The trend factor is measured taking into account the dynamics of speech. Evaluation of the proposed algorithm and comparisons with state of the art systems will be performed in the context of varying automobile noises.

Method for temporal interpolation of short-term spectra and its application to adaptive system identification

In this paper an extension of analysis filterbanks utilized in adaptive system identification schemes is presented. The extension operates as a postprocessing stage following conventional polyphase filterbanks or short-term Fourier transforms. The idea of the new method is to exploit the redundancy of succeeding short-term spectra for computing interpolated temporal supporting points. For its efficient implementation some approximations are performed - it can be shown that the postprocessing stage can be easily realized based on the weighted sum of subband signals. The new method allows for significant increase of the frameshift (subsampling rate), leading to a reduction of the computational complexity while keeping the convergence speed and the steady-state performance. Alternatively, the frameshift can be kept unchanged - in this case an improved steady-state convergence can be achieved. Real-time measurements performed with systems for acoustic echo cancellation have shown that significant improvements in terms of echo reduction can be achieved while increasing the amount of required memory only marginally.

Acoustic Array Processing for Speech Enhancement

This chapter contains sections titled: Introduction Signal Processing in Subband Domain Multichannel Echo Cancellation Speaker Localization Beamforming Sensor Calibration Postprocessing Conclusions References

Model-based speech enhancement for automotive applications

In this contribution we present a new approach for speech signal enhancement that improve a distorted speech signal in low SNR scenarios (at least if only a portion of the frequency range is highly disturbed). The approach consists of three algorithmic parts: a standard noise suppression unit, a partial speech reconstruction unit, and a time-frequency selective mixing unit. The basic idea of partial speech reconstruction is first to detect time-frequency areas where the SNR is still acceptable. These areas are used to extract relevant signal features such as the pitch frequency or the spectral envelope. Based on pre-trained signal models and the extracted features, the signal is reconstructed in a second stage. Finally, the reconstructed and the conventionally enhanced signals are mixed in a time-frequency selective manner. Subjective and objective tests indicate that a significant quality improvement is possible compared to conventional schemes - especially in high noise conditions.

Signal processing techniques for seat belt microphone arrays

Microphones integrated on a seat belt are an interesting alternative to conventional sensor positions used for hands-free telephony or speech dialog systems in automobile environments. In the setup presented in this contribution, the seat belt consists of three microphones which usually lay around the shoulder and chest of a sitting passenger. The main benefit of belt microphones is the small distance from the talker’s mouth to the sensor. As a consequence, an improved signal quality in terms of a better signal-to-noise ratio (SNR) compared to other sensor positions, e.g., at the rear view mirror, the steering wheel, or the center console, can be achieved. However, the belt microphone arrangement varies considerably due to movements of the passenger and depends on the size of the passenger. Furthermore, additional noise sources arise for seat belt microphones: they can easily be touched, e.g., by clothes, or might be in the path of an air-stream from the automotive ventilation system. This contribution presents several robust signal enhancement algorithms designed for belt microphones in multi-seat scenarios. The belt microphone with the highest SNR (usually closest to the speaker’s mouth) is selected for speech signal enhancement. Further improvements can be achieved if all belt microphone signals are combined to a single output signal. The proposed signal enhancement system for belt microphones includes a robust echo cancelation scheme, three different microphone combining approaches, a sophisticated noise estimation scheme to track stationary as well as non-stationary noise, and a speech mixer to combine the signals from each seat belt to a single channel output in a multi-seat scenario.

Refinement and Temporal Interpolation of Short-Term Spectra: Theory and Applications

In this contribution, methods for spectral refinement (SR) and spectral interpolation (SI) are presented. These methods can be implemented as a post-processing stage after conventional frequency analyses such as overlap add-based decomposition schemes. The principle idea of SR is to individually refine each subband signal after frequency decomposition and to compute additional frequency-supporting points in between using a linear combination of a few neighboring (in terms of time and frequency) subband signals. For efficient implementation, a simplification of the SR method has been derived—it has been shown that the refinement can easily be implemented using short FIR filters in each subband, resulting in a very low computational complexity. The SI method exploits the redundancy of succeeding short-term spectra for computing interpolated temporal supporting points in between the originally generated frames. This is achieved by efficient approximations, and the whole method can be realized on weighted sums of subband signals. The new interpolation method can be applied in adaptive system identification schemes (e.g., echo cancellation or channel estimation), allowing for a significant increase of the frameshift (subsampling rate). This leads to a reduction of the computational complexity while keeping the convergence speed and the steady-state performance constant. Alternatively, the frameshift can be kept the same. In this case an improved steady-state convergence can be achieved. The proposed method for SR has been applied as a preprocessing stage for fundamental frequency (pitch frequency) estimation, and the SI method has been utilized for subband echo cancellation. Evaluations have shown that pitch frequency estimation can be improved significantly for all considered signal-to-noise ratios when employing the SR method. Real-time measurements performed on systems for acoustic echo cancellation have demonstrated that significant improvements in terms of echo reduction can be achieved while only marginally increasing the amount of required memory.

Chapter 8. Acoustic Array Processing for Speech Enhancement

This chapter contains sections titled: Introduction Signal Processing in Subband Domain Multichannel Echo Cancellation Speaker Localization Beamforming Sensor Calibration Postprocessing Conclusions References