Search

This paper reports our work on building up a Cantonese Speech-to-Text (STT) system with a syllable based acoustic model. This is a part of an effort in building a STT system to aid dyslexic students who have cognitive deficiency in writing skills but have no problem expressing their ideas through speech. For Cantonese speech recognition, the basic unit of acoustic models can either be the conventional Initial-Final (IF) syllables, or the Onset-Nucleus-Coda (ONC) syllables where finals are further split into nucleus and coda to reflect the intra-syllable variations in Cantonese. By using the Kaldi toolkit, our system is trained using the stochastic gradient descent optimization model with the aid of GPUs for the hybrid Deep Neural Network and Hidden Markov Model (DNN-HMM) with and without I-vector based speaker adaptive training technique. The input features of the same Gaussian Mixture Model with speaker adaptive training (GMM-SAT) to DNN are used in all cases. Experiments show that the ONC-based syllable acoustic modeling with I-vector based DNN-HMM achieves the best performance with the word error rate (WER) of 9.66% and the real time factor (RTF) of 1.38812., Comment: 7 pages, 3 figures, LREC 2016

This paper considers the reconstruction of a sparse coefficient vector {\theta} for a rational transfer function, under a pair of FIR and Takenaka-Malmquist (TM) bases and from a limited number of linear frequency-domain measurements. We propose to concatenate a limited number of FIR and TM basis functions in the representation of the transfer function, and prove the uniqueness of the sparse representation defined in the infinite dimensional function space with pairs of FIR and TM bases. The sufficient condition is given for replacing the l_0 optimal solution by the l_1 optimal solution using FIR and TM bases with random samples on the upper unit circle, as the foundation of reconstruction. The simulations verify that l_1 minimization can reconstruct the coefficient vector {\theta} with high probability. It is shown that the concatenated FIR and TM bases give a much sparser representation, with much lower reconstruction order than using only FIR basis functions and less dependency on the knowledge of the true system poles than using only TM basis functions.

Most rational systems can be described in terms of orthonormal basis functions. This paper considers the reconstruction of a sparse coefficient vector for a rational transfer function under a pair of orthonormal rational function bases and from a limited number of linear frequency-domain measurements. We prove the uncertainty principle concerning pairs of compressible representation of orthonormal rational functions in the infinite dimensional function space. The uniqueness of compressible representation using such pairs is provided as a direct consequence of uncertainty principle. The bound of the number of measurements which guarantees the replacement of 1_0 optimization searching for the unique sparse reconstruction by 1_1 optimization using random sampling on the unit circle with high probability is provided as well.