Roman A. Dyba

Partial update PNLMS algorithm for network echo cancellation

The Proportionate Normalized Least Mean Square (PNLMS) algorithm has been proposed for network echo cancellation to take advantages of the sparseness of the echo path impulse responses. The PNLMS algorithm has fast initial convergence but slows down dramatically after the initial period. In this paper, a novel algorithm to combine the PNLMS algorithm with the technique of partial update is proposed. Simulation results show that the proposed algorithm can achieve faster overall convergence with less computation.

Efficient partial update algorithm based on coefficient block for sparse impulse response identification

Network echo path impulse responses with large echo coverage capacity are sparse. Two coefficient-based sparse partial update adaptive algorithms are described in the work of Deng and Doroslovacki (2004) to achieve fast convergence by taking advantages of the sparseness of the echo path and the partial update technique. In this paper, a novel block sparse partial update normalized least mean square (BSPNLMS) algorithm is proposed. Simulation results show that the proposed algorithm can have better convergence performance than the coefficient-based algorithms with much less computational complexity.

Parallel structures for fast estimation of echo path pure delay and their applications to sparse echo cancellers

The networking environment allowing for coexistence of voice and data communication has become complex, and voice aspects of telecommunication, particularly in voice-over-IP networks, demand echo cancellers to cover all voice channels, as opposed to only long-haul channels, as it used to be in traditional public switched telephone networks. Echo path coverage requirements for the echo cancellers have become more demanding, which contributed to an increase of computational cost of their implementations. One of the methods of decreasing that cost is via taking advantage of sparseness of the echo path impulse response (sparse systems). In addition to requirements related to reducing an overall computational cost, it is desirable to accelerate the adaptation, so the voice quality is improved. In the case of sparse systems, the requirement of accelerating the adaptation translates into faster allocation of the main adaptive filter window. This study explores an approach that takes advantage of the relation between the FIR filter length and the convergence speed, with focusing of the subrate adaptive filter that is being used for identifying the pure delay. The proposed approach is suitable for single reflectors. The proposed idea can be expanded to cover multi-reflectors. The presented results of numerical experiments relate to the NLMS. Yet the approach is general and can be applied to other adaptive filter algorithms. The computational overhead associated with the parallel structure (and related decision algorithm) is tangible yet quite minor. This additional cost is split over two distinct parts: (a) filter window overlap to ensure the echo path delays that fit to the parallel structure partitioning are treated without disadvantage and (b) the additional decision logic that is used to identify the position (within one of the M parallel branches covering approximately 1/Mth of the entire echo path coverage each) of the echo path impulse response peak. The principal benefit of the parallel structure is the adaptation speed increase (and this increase is a function of M) at an expense of only very minor computational cost.

Efficient Implementation of Voice Enhancement Algorithms on Freescale Starcore SC3400 Based Digital Signal Processor

Efficiently implementing complex voice enhancement algorithms on digital signal processor platforms is a challenging task. A wide scope of knowledge and skills to understand algorithm details, DSP architecture and instruction set, fractional number arithmetic, performance measurement and optimization are essential. In this paper, implementation aspects of the typical voice enhancement related algorithms on Freescale StarCore SC3400 based DSP platform are presented.

Fixed-Point Implementation of Noise Reduction Using StarCore- SC3400

The presence of relatively high-level background noise in a telecommunication channel may lower the perceived voice quality of speech signals as well as degrade in-band signaling. The challenge is to reduce the noise to a satisfactory level while minimizing the use of computational resources. This paper describes an efficient noise reduction algorithm and its implementation on a high-performance Digital Signal Processor (DSP) based on the Freescale StarCore SC3400 core. The NR implementation methodology takes advantage of the StarCore architecture and Code Warrior development tools to reduce engineering efforts. After performing code optimization by exploring a mix of high-level and machine-level languages, the NR computational cost is reduced to approximately 0.6 Millions of Cycles Per Second for the narrow-band applications (i.e., G. 711 with 8kHz sampling rate). The algorithm has been evaluated using different approaches, including subjective evaluation by expert listeners. An overall noise reduction of 10-12dB (for standard setting of 13dB noise reduction threshold) has been achieved for most natural-speech signals polluted with stationary noise. The noise reduction component has also been evaluated using wideband signals (16kHz sampling rate). The machine cycle count increased to 1 MCSP (approximately) while the overall noise reduction of 9-12dB was achieved without observing adverse side effects related to voice quality.

Development of DSP Based Network Echo Cancellers

Adaptive filtering is an important area of digital signal processing and is one of the main topics in the digital signal processing classrooms. Echo cancellation is a typical example used to introduce adaptive filtering algorithms. In this paper, filling the gap between the understanding of adaptive filtering theory and the knowledge and skills needed to build a practical echo canceller is attempted. The necessary knowledge and skills (such as algorithm design, implementation on DSP platforms, debugging, performance analysis and testing) for developing a high-performance network echo cancellers are covered.

Fast and efficient proportionate adaptive algorithms for pure delay estimation

Pure delay estimator is a vital component in network echo cancellation systems targeting sparse echo paths with unknown pure delays. In this paper, the Proportionate Normalized Least Mean Square (PNLMS) algorithm is applied to a sub-rate pure delay estimator. Compared with the Normalized Least Mean Square (NLMS) algorithm [1], the PNLMS algorithm can detect the pure delay faster but the increased computational complexity is high. To make the overall system more efficient for implementing on DSP platforms, two novel simplified proportionate adaptive algorithms are proposed. The simulation results demonstrate that the proposed algorithms offer comparable performance of pure delay estimation as the PNLMS algorithm with much less computational complexity.

Identification of Multi-Reflective Echo Paths using Sub-Rate Adaptive Filtering in Presence of Moderate Aliasing

Conditions for coexistence of voice and data communication have become complex, and voice aspects of telecommunication, particularly in Voice-over-IP networks, demand echo cancellers to cover all voice channels, as opposed to only long-haul channels, as it used to be in traditional public switched telephone networks. Echo path coverage requirements for the echo cancellers have become more demanding, causing an increase of computational cost of their implementations. One of the methods of decreasing that cost is via implementing sparse echo cancellers. This study explores an approach to additionally reduce computational cost at the stage of pre-processing of input signals for use in a sub-rate adaptive filter, at the expense of aliasing effects. This reduction is based on observations indicating that adaptive filters are not very sensitive to aliasing effects when it comes to detection of multiple reflections.