文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > IT/计算机 > 互联网 > MULTI-LAYER F0 MODELING FOR HMM-BASED SPEECH SYNTHESIS
侵权投诉

MULTI-LAYER F0 MODELING FOR HMM-BASED SPEECH SYNTHESIS

MULTI-LAYER F0 MODELING FOR HMM-BASED SPEECH SYNTHESIS

Cheng-Cheng Wang, Zhen-Hua Ling, Bu-Fan Zhang, Li-Rong Dai iFlytek Speech Lab, University of Science and Technology of China, Hefei

ABSTRACT

This paper proposes a two-layer fundamental frequency (F0) modeling method for HMM-based parametric speech synthesis. The F0 models are trained for each context-dependent phoneme in the conventional HMM-based speech synthesis system. Considering the super-segmental characteristics of F0 features, an explicit syllable-layer F0 model is introduced in this paper. At synthesis stage, the F0 contour is generated by maximizing the combined likelihood functions of the phone-layer and syllable-layer F0 models. The objective and subjective evaluation results in our experiments show that the proposed multi-layer F0 modeling method can improve the performance of F0 prediction for emotional speech synthesis.

Index Terms—Speech synthesis, hidden Markov model, v model, fundamental frequency modeling

1. INTRODUCTION

lied to sp

ctrum f

ture. P

phone, while the contour of F0 beyond the phone layer. The F0 contour is associated with some high-level prosodic features with larger scope than a phone, such as tones of a syllable, stress in a phrase, prosodic boundary in a sentence, etc. Further more, emotional speech is influenced by not only the contextual information but also some paralinguistic features. It is difficult to study and model the paralinguistic variations from only the F0 features within a phone. Therefore, the construction of long-term F0 models is of important significance.

Furthermore, syllable is treated as the basic unit for F0 modeling by many prosodic models, such as target model [1], Fujisaki model [2] and SFC model [3]. In these models, the super-segmental characteristics of F0 are modeled explicitly. Take the SFC model for example, the F0 contour is assumed to be a superposition of several tiers, including

clause tier, phrase tier, syllable tier and so on, and syllable is the smallest unit for modeling.

Based on the discussions above, a new F0 modeling method which combines syllable-layer model and phone-layer model for HMM-based speech synthesis is proposed. We can get much long-term information about F0 features by means of syllable model. Experimental results show that is impl

n accu

peech dat

2. METHOD

2.1. Model Training 2.1

The overall training and synthesis processes of HMM-based speech system is shown in Fig.1. In this paper, we only focus on the construction of F0 model. The spectrum part is the same as the conventional one [4]. Acoustic features are extracted from the speech waveforms of training database. STRAIGHT [5] as a high quality speech vocoder is adopted here to analyze the spectral envelop and F0 for each speech frame. In context-dependent phone model training, the feature for each frame consists of static, delta and delta-delta components of logarithmized F0. The context- dependent HMMs are estimated according to the acoustic features and context labels under the maximum likelihood criterion [6]. F0 stream is modeled by a multi-space probability distribution (MSD) [7]. In our experiments, we use two-space probability distribution, which represent the either of voiced and unvoiced frames. A decision tree based model clustering method is applied to deal with data-sparsity problems and predict the context-dependent models outside the training set. In training period, the mean and variance parameters of F0 and spectrum features are estimated for each context-dependent phone HMM [4].

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email