Application of neural networks to articulatory motion estimation

Abstract

A B S T R A C TThe first five parameters the shape of the oral and pha This paper discusses an application of neural networks to the problem of estimating the motion of articulatory organs from speech waves.Recently, neural networks (NN) have been studied extensively. It has been proved that a three-or four-layer feed-forwardIn this paper, we apply this feature of NN to the articulatory parameter estimation problem.The evaluation test is performed using the vowel data in 5201) tokens in the ATR word database. Our results show that the difference in estimated articulatory parameter values between the conventional method (MM) and NN is only 0.1, which is about 3 % of value range, on average. For a few data, big differences arise between MM and NN, but this is due to mis-estimation in MM rather than NN. The percentage of misestimates in NN is less than 50 % of MM. As for calculation time, NN is 10 times faster than MM.Thus, a high speed and stable articulatory parameter estimation technique is realized using neural networks. latory For that purpose, the neural networks can approximate arbitrary nodinear functions. I. I N T R O D U C T I O NIt is well known that the coarticulation compensation and speaker adaptation are two major difficulties in speech recognition. These problems might be considered most clearly and fundamentally from the view point of the speech production mechanism. On the basis of this idea, the model matching (MM) method was proposed to extract articulatory parameters, which represent articulatory movements, from speech waves [1,2] and the estimated parameters were used for speech recognition [3,4,5]. These parameters are effective for the above two problems but the conventional estimation method has some problems: one is calculation cost and the other is instability of estimation. Since this method is constructed on the basis of hill climbing methods, it requires many iterations to converge and sometimes finds only local minimum. For a speech recognition system, a faster and more stable estima tion method is desired.Recently, neural networks (NN) have been studied extensively [7]. It has been proved that a three or four-layer feedforward neural networks can approximate arbitrary nonlinear functions [b]. In this paper, we apply this feature of NN to the articulatory parameter estimation problem. Fig.1 Articulatory m A R T I C U L A T O R Y M O D E LThe total configuration of the model and the characteristics of the articulatory parameters are shown in Fig.1 and Table 1.-489 -

Retrieving Tract Variables From Acoustics: A Comparison of Different Machine Learning Strategies

“…Shirai et al [101] proposed an analysis-by-synthesis approach, which they termed as “Model Matching,” where speech was analyzed to generate articulatory information and then the output was processed by a speech synthesizer such that it had minimal distance from the actual speech signal in the spectral domain. Kobayashi et al [64] proposed a feed-forward MLP architecture with two hidden layers that uses the same data as used in [101] to predict the articulatory parameters and showed faster performance and better estimation accuracy. Regression techniques have been explored a number of times for speech inversion.…”

Acoustic-Articulatory Modeling With the Trajectory HMM

“…Previous attempts to recover articulatory movement from the speech signal involved building a mapping from the acoustic domain to the articulatory domain, either manually or constructed automatically from parallel data [4], [5], [6], [7], [8], [9], [10], [11], [12]. Variations of neural networks [5], [13], [6], [11] have become popular in the latter category. Often the inversion system is built separately from the recognition framework, particularly because the slowly varying nature of articulation may be best modelled in a different way to speech acoustics which change more rapidly, and are noisier.…”

Minimizing speaker variation effects for speaker-independent speech recognition

“…Nonlinear mapping of two different observation spaces is of great interest for both theoretical and practical purposes. In the area of speech processing, nonlinear mapping has been applied to noise enhancement [1,32], articulatory motion estimation [29,18], and speech recognition [16]. Neural networks have been used successfully to transform data of a new speaker to a reference speaker for speaker-adaptive speech recognition [11].…”