my papers

now total is 4

NO.1

A New Markov Model for Speech Recognition

Zhu Qifeng , 14th Lab of the Institute of Acoustic, 11.11,1995 [ Abstract ] In this paper we give out and emphasize the concept of the nonlinear property of the information distribution of speech signal in the time domain. Some exist solutions are reviewed. With two pragmatic principles we give out a new method to modify the traditional HMM to fit this property. This method is tested in non- continuous speech recognition. The result is also reported in this paper and discussed. Key Word: HMM, speech recognition, nonlinear Topic area: (B) Speech Recognition Mail address: 14 Lab , mailbox 2712, Beijing, 100080 Email: zqf@speech1.ioa.ac.cn Telephone: 2553842

NO.2.

AMM algorism for speech recognition

Zhu Qifeng, 14th Lab of the  Institute of  Acoustic

Abstract
        In this paper I first discuss the basis of the HMM 
model. In HMM model the training and the matching algorithm 
are with the goal to maximize the probability. From the view of 
pattern recognition and robust training, the goal should be 
maximize the distance among samples in different group and 
to minimize the distance among samples in a same group. In 
this paper we will find the basis of HMM is not the best to fit 
the goal of pattern recognition, and maybe is not the most 
robust one in some occurrences.
        In the second half of this paper, I will discuss a 
straitforward way to determine the states in the HMM model, 
using the concept of  "sound stimuli" initiated by Prof. Yu. 
The training process is greatly simplified and no recursive process 
is needed. I discuss the potential of this method to fit the goal 
of pattern recognition.

Key word: HMM, speech recognition

NO.3

The Gain Factor In LPC and CEP coefficients In the Speech Recognition

Zhu Qifeng , 14th Lab of the Institute of Acoustic, 4,15,1996 [ Abstract ] In the LPC analysis in speech recognition, the gain factor, G, and the C0 in CEP coresponds with the simulation of vocal cords, it is firmly related with the voice amplitude. In this paper, we will illustrate the meaning of the gain factor by means of many figures, including the figures of LPC spectrum, CEP spectrum, frequency spectrum derived from LPC, each of which we give a figure with and without the gain factor. And we also give out the origin frequency spectrum of the voice. After this, we also calculate the standard deviation of the CEP coefficients of different orders. In this way, the weighting value can be obtained. With the direct means and theoretical means above mentioned, we suggest that the gain factor can be omitted.

NO.4

The coordinated hearing phenomena and its influnce in speech recognition

Zhu Qifeng , 14th Lab of the Institute of Acoustic, 5,15,1996 [ Abstract ] In this paper, we would discuss the phenomena of coodinated hearing, i.e. the sense of a voice at one time is not only depended on the frequency spectrum at the current time, but also depended on the frequency spectrum of past of the future sense. After the introduction, we would give out some examples, by which the coordinated phenomena we have studied can be clearly seen. Then with the method of blackbox, we would give preliminary modelized explanation for the phenomena. At the end, we would discuss the effect of these phenomena on the speech recognition, and give out some discuss on the possible improvement on the algorithm on speech recognition and the HMM model. Key Word speech recognition, hearing£¬ coordinated hearing

ok?

back home

For the whole paper, please contact me with zqf@ucla.edu