Contextual biasing is an important and challenging task for end-to-end automatic speech recognition (ASR) systems, which aims to achieve better recognition performance by biasing the ASR system to particular context phrases such as person names, music list, proper nouns, etc. Existing methods mainly include contextual LM biasing and adding bias encoder into end-to-end ASR models. In this work, we introduce a novel approach to do contextual biasing by adding a contextual spelling correction model on top of the end-to-end ASR system. We incorporate contextual information into a sequence-to-sequence spelling correction model with a shared context encoder. The proposed model includes two different mechanisms: autoregressive (AR) and non-autoregressive (NAR). We also propose filtering algorithms to handle large-size context lists, and performance balancing mechanisms to control the biasing degree of the model. The proposed model is a general biasing solution which is domain-insensitive and can be adopted in different scenarios. Experiments show that the proposed method achieves as much as 51% relative word error rate (WER) reduction over ASR system and outperforms traditional biasing methods. Compared to the AR solution, the NAR model reduces model size by 43.2% and speeds up inference by 2.1 times.

The design of predictive quantizers generally suffers from difficulties due to the prediction loop, which have an impact on the convergence and the stability of the design procedure. We previously proposed an asymptotically closed-loop approach to quantizer design for predictive coding applications, which benefits from the stability of open-loop design while asymptotically optimizing the actual closed-loop system. In this paper, we present an enhancement to the approach where joint optimization of both predictor and quantizer is performed within the asymptotically closed-loop framework. The proposed design method is tested on synthetic sources (first-order Gauss and Laplacian-Markov sequences), and on natural sources, in particular, line spectral frequency parameters of speech signals.

This correspondence builds on the asymptotic closed-loop approach to predictive vector quantizer design (Khalil et al. 2001), and extends it to the design of predictive multistage vector quantizers for low bit rate video coding. The design approach resolves longstanding shortcomings, in particular, design stability and empty-cell problems. Simulation results show substantial gains over traditional design approaches.

The basic vector quantization (VQ) technique employed in video coding belongs to the category of predictive vector quantization (PVQ), as it involves quantization of the (motion compensated) frame prediction error. It is well known that the design of PVQ suffers from fundamental difficulties, due to the prediction loop, which have an impact on the convergence and the stability of the design procedure. We propose an approach to PVQ design that enjoys the stability of open-loop design while it ensures ultimate optimization of the closed-loop system. The method is derived for general predictive quantization, and we demonstrate it on video compression at low bit rates, where it provides substantial improvement over standard open and closed loop design techniques. Further, the approach outperforms standard DCT-based video coding.

The design of a multi-stage vector quantizer (MSVQ) based source-channel coding system is optimized for packet networks. Resilience to packet loss is further enhanced by a proposed interleaving approach that ensures that a single lost packet only eliminates a subset of the vector stages. The design is optimized while taking into account compression efficiency, packet loss rate, and the interleaving technique in use. The new source-channel-optimized MSVQ is tested on memoryless line spectral frequency (LSF) parameter quantization in speech coders. A source-channel-optimization MSVQ is shown to yield a gain of up to 2.0 dB in SNR, for coding the LSFs, over traditional MSVQ and to substantially enhance the robustness of packetized speech transmission. Although the formulation is given in the context of packet networks, the work is directly extendible to the broader category of erasure channels.

The design of a source-channel system based on the multi-stage vector quantizer (MSVQ) is optimized for packet networks. Resilience to packet loss is further enhanced by a proposed interleaving scheme that ensures that a single lost packet only eliminates a subset of the vector stages. The design is optimized while taking into account compression efficiency, packet loss rate, and the interleaving technique in use. Simulation results demonstrate that the packet network-optimized MSVQ outperforms traditional MSVQ for a variety of images and channel conditions, with gains of up to 2.0 dB in PSNR. Although the formulation is given in the context of packet networks, the work is directly extendible to the broader category of erasure channels.

This work considers the design of a multi-stage vector quantizer (MSVQ) for application to the motion compensated prediction error of video signals. It is well known that the design of predictive vector quantizers suffers from fundamental difficulties due to the prediction loop, which have an impact on the convergence and the stability of the design procedure. We propose an approach to predictive MSVQ design that enjoys the stability of open-loop design while ensuring ultimate optimization of the closed-loop system. The proposed design method is tested on video compression at low bit rates, where it significantly outperforms widely-used closed-loop design techniques, and achieves improvement over the H.263 standard.

Traditional wavelet packet (WP) optimization techniques neglect information about the structure of the lossy part of the compression scheme. Such information, however, can help guide the optimization procedure so as to result in efficient WP structures. We propose a wavelet packet algorithm with a constrained rate-distortion optimization which makes it suited to subsequent tree-structured coding such as with the set partitioning in hierarchical trees (SPMT) algorithm. The (octave-band) wavelet transform lends itself to simple and coherent tree-shaped spatial relations which can then be used to define zero-trees. Yet, input images have different frequency distributions and an adaptive transform such as WP is bound to be more efficient on an image-by-image basis. With WP algorithms, the coefficients in the WP domain can be rearranged to produce what resembles (or simulates) the normal wavelet transform structure. This stage is usually performed to simplify the coding stage. However, an unconstrained optimization can result in a transformed image with complicated or incoherent tree-shaped spatial relations. This work aims to show that the efficiency of embedded coders such as SPMT and Shapiro's Zerotrees strongly depends on WP structures with coherent spatial tree relationships.

In this paper, we present a 3-D wavelet compression technique for image sequences in video conferencing applications. One of the main requirements for such applications is that the delay has to be within some acceptable limit. When applying wavelet decomposition in the temporal direction, we must store a large number of frames so that such a decomposition can be effective. But this translates into a correspondingly large algorithmic delay. A technique to avoid this obstacle will be proposed. The basic idea is to overlap the decomposition with previous frames that have already been transmitted and selectively transmitting only the new wavelet coefficients.

Three-dimensional (3-D) video compression using wavelets decomposition along the temporal axis dictates that a number of video frames must be buffered to allow for the temporal decomposition. Buffering of frames allows for the temporal correlation to be made use of, and the larger the buffer the more effective the decomposition. One problem inherent in such a set up in interactive applications such as video conferencing, is that buffering translates into a corresponding time delay. We show that 3-D coding of such image sequences can be achieved in the true sense of temporal direction decomposition but with much less buffering requirements. For a practical coder, this can be achieved by introducing an approximation to the way the transform coefficients are encoded. Applying wavelet decomposition using some types of filters may introduce edge errors, which become more prominent in short signal segments. We also present a solution to this problem for the Daubechies (1988) family of filters.

A system is developed for the compression of video images, with particular emphasis on video conferencing applications. It combines the well-known quadtree decomposition approach and the discrete cosine transform in an adaptively spatial subsampling and motion compensated scheme. A comparison with the ITU-T H.261 standard for low bit rate video coding reveals that it has better SNR, better perceptual quality (less blocking artifacts), less computational load (about four times less), as well as ease of hardware implementation.