Sung-Hsien Hsieh

A real-time Mandarin dictation machine for Chinese language with unlimited texts and very large vocabulary

A successfully implemented real-time Mandarin dictation machine which recognizes Mandarin speech with unlimited texts and very large vocabulary for the input of Chinese characters to computers is described. Isolated syllables including the tones are first recognized using specially trained hidden Markov models with special feature parameters. The exact characters are then identified from the syllables using a Markov Chinese language model. The real-time implementation is on an IBM PC/AT, connected to a set of special hardware boards on which ten TMS 320C25 chips operate in parallel. It takes only 0.45 s to dictate a character.<<ETX>>

Compressed Sensing Detector Design for Space Shift Keying in MIMO Systems

Space shift keying (SSK) modulation and its extension, the generalized SSK (GSSK), present an attractive framework for the emerging large-scale MIMO systems in reducing hardware costs. In SSK, the maximum likelihood (ML) detector incurs considerable computational complexities. We propose a compressed sensing based detector, NCS, by formulating the SSK-type detection criterion as a convex optimization problem. The proposed NCS requires only O(ntNrNt) complexity, outperforming the O(NrNtnt) complexity in the ML detector, at the cost of slight fidelity degradation. Simulations are conducted to substantiate the analytical derivation and the detection accuracy.

Special speech recognition approaches for the highly confusing Mandarin syllables based on hidden Markov models

Abstract In this paper, several special speech recognition approaches based on hidden Markov models (HMMs) are presented for the highly confusing Mandarin syllables by considering the characteristics of the vocabulary. This is because there are totally 408 syllables (disregarding the tones) in Mandarin speech, and it is believed that correct recognition of these syllables is the key to the development of a Mandarin dictation machine which recognizes Mandarin speech with very large vocabulary and unlimited texts. However, accurate recognition of these 408 syllables is very difficult because there exist 38 confusing sets among them, each of which has at most 19 very confusing syllables. Direct application of conventional standard approaches of HMMs to these syllables gives recognition rates in the order of only 70–80%, thus several special approaches are proposed in this paper to provide better performance. Some of these approaches concentrate on the training algorithms for the HMMs, including the two-pass training, the revised two-pass training, the three-pass training, and the revised three-pass training approaches. The basic idea is to protrude the very short initial parts (initial consonant parts) and de-emphasize the final parts (vowel or diphthong parts but including possible medial or nasal ending) of the syllables such that the confusing syllables can be better distinguished. Also, some other approaches concentrate on the recognition phase of the HMMs, including the state duration bounds and the two-stage search strategy, which can also improve the recognition rates and/or speeds. A special hardware is also implemented to complete all the recognition operations in real-time, on which all the approaches discussed in this paper can be applied.

Sparse Fast Fourier Transform by downsampling

Sparse Fast Fourier Transform (sFFT) [1][2], has been recently proposed to outperform FFT in reducing computational complexity. Assume that an input signal of length N in the frequency domain is K-sparse, where K ≤ N. sFFT costs O(K logN) instead of O(N logN) in FFT. In this paper, a new fast sFFT algorithm is proposed and costs O(K logK) averagely without any operations being related to N. The idea is to downsample the original input signal at the beginning. Subsequent processing operates under downsampled signals, which length is proportional to O(K). However, downsampling possibly leads to “aliasing.” By shift theorem of DFT, the aliasing problem can be formulated as the “Moment-preserving problem.” In addition, a top-down iterative strategy combined with different downsampling factors further saves computational costs. Complexity analysis and experimental results show that our method outperforms FFT and sFFT.

2D sparse dictionary learning via tensor decomposition

The existing dictionary learning methods mostly focus on ID signals, leading to the disadvantage of incurring overload of memory and computation if the size of training samples is large enough. Recently, 2D dictionary learning paradigm has been validated to save massive memory usage, especially for large-scale problems. To address this issue, we propose novel 2D dictionary learning algorithms based on tensors in this paper. Our learning problem is efficiently solved by CANDECOMP/PARAFAC (CP) decomposition. In addition, our algorithms guarantee sparsity constraint, which makes that sparse representation of the learned dictionary is equivalent to the ground truth. Experimental results confirm the effectness of our methods.

Phase transition of joint-sparse recovery from multiple measurements via convex optimization

In sparse signal recovery of compressive sensing, the phase transition determines the edge, which separates successful recovery and failed recovery. Moreover, the width of phase transition determines the vague region, where sparse recovery is achieved in a probabilistic manner. Earlier works on phase transition analysis in either single measurement vector (SMV) or multiple measurement vectors (MMVs) is too strict or ideal to be satisfied in real world. Recently, phase transition analysis based on conic geometry has been found to close the gap between theoretical analysis and practical recovery result for SMV. In this paper, we explore a rigorous analysis on phase transition of MMVs. Such an extension is not intuitive at all since we need to redefine the null space and descent cone, and evaluate the statistical dimension for ℓ2,1-norm. By presenting the necessary and sufficient condition of successful recovery from MMVs, we can have a boundary on the probability that the solution of a MMVs recovery problem by convex programming is successful or not. Our theoretical analysis is verified to accurately predict the practical phase transition diagram of MMVs.

Fast OMP: Reformulating OMP via iteratively refining ℓ 2 -norm solutions

Orthogonal matching pursuit (OMP) is a powerful greedy algorithm in compressed sensing for recovering sparse signals despite its high computational cost for solving large scale problems. Moreover, its theoretic performance analysis based on mutual incoherence property (MIP) is still not accurate enough. To overcome these difficulties, this paper proposes a fast OMP (FOMP) algorithm by reformulating OMP in terms of refining ℓ2-norm solutions in a greedy manner. ℓ2-norm solutions are known for being non-sparse, but we show that the ℓ2-norm solution associated with a greedy structure actually solves the sparse signal reconstruction problem well. We analyze exact recovery of FOMP via an order statistics probabilistic model and provide practical performance bounds.

A fully parallel Mandarin speech recognition system with very large vocabulary and almost unlimited texts

The authors describe a fully parallel real-time Mandarin dictation machine which recognizes Mandarin speech with almost unlimited texts and a very large vocabulary for the input of Chinese characters to computers. Isolated syllables including the tones are first recognized using specially trained hidden Markov models with special feature parameters, and the exact characters are then identified from the syllables using a Markov Chinese language model, because every syllable can represent many different homonym characters. The real-time implementation is in Occam language on a transputer system with 10 T800 processors operating in parallel. The overall correction rate for the final output characters is about 80%.<<ETX>>

Fast binary embedding via circulant downsampled matrix

Binary embedding of high-dimensional data aims to produce low-dimensional binary codes while preserving discriminative power. State-of-the-art methods often suffer from high computation and storage costs. We present a simple and fast embedding scheme by first downsampling N-dimensional data into M-dimensional data and then multiplying the data with an M×M circulant matrix. Our method requires O(N + M log M) computation and O(N) storage costs. We prove if data have sparsity, our scheme can achieve similarity-preserving well. Experiments further demonstrate that though our method is cost-effective and fast, it still achieves comparable performance in image applications.

Privacy-preserving data collection and recovery of compressive sensing

Energy-efficient data collection and privacy-preserving data recovery have received much attention recently. We propose the first encryption framework for the computation-intensive basis pursuit problem to be securely solved in the cloud with the data being efficiently collected using compressive sensing. We provide security and efficiency analyses to show the effectiveness of our method. Simulations and comparison with state-of-the-art are also conducted.