Compress Polyphone Pronunciation Prediction Model with Shared Labels

Compress Polyphone Pronunciation Prediction Model with Shared Labels Pengfei Chen, Lina Wang, Hui Di, Kazushige Ouchi and Lvhong Wang Research & Development Center, Toshiba, China {chenpengfei, wanglina, dihui}@toshiba.com.cn kazushige.ouchi@toshiba.co.jp wlv1990@gmail.com Abstract It is well known that deep learning model has huge parameters and is computationally expensive, especially for embedded and mobile devices. Polyphone pronunciations selection is a basic func- tion for Chinese Text-to-Speech (TTS) application. Recurrent neural network (RNN) is a good sequence labeling solution for polyphone pronunciation selection. However, huge parameters and computation make compression needed to alleviate its disadvantages. Meanwhile, Large- scale-labels classification leads to more complicated network and heavy computation cost. In contrast to existing quantization with low precision data format and projection layer, we propose a novel method based on shared labels, which focuses on compressing the fully-connected layer before Softmax for models with a huge number of labels in TTS polyphone selection. The basic idea is to compress large number of target labels into a few label clusters, which will share the parameters of fully-connected layer. Furthermore, we combine it with other methods to further compress the polyphone pronunciation selection model. The experimental result shows that for Bi-LSTM (Bidirectional Long Short Term Memory) based polyphone selection, shared labels model decreases about 52% of original model size and accelerates prediction by 44% almost without performance loss. It is worth mentioning that the proposed method can be applied for other tasks to compress model and accelerate calculation. Keywords: Bi-LSTM, Polyphone Pronunciation Prediction, Model Compression, Shared La- bels. 1 Introduction Polyphone pronunciation prediction is a basic module of G2P (Grapheme-to-Phoneme) in Chinese Text- to-Speech (TTS) system, which provides the right pronunciation for Chinese character. The algorithms of polyphone pronunciation prediction include dictionary-based algorithm, statistical machine learning- based algorithm like Conditional Random Field (CRF) in (Lafferty et al., 2001), and deep learning-based algorithm like RNN. Dictionary-based method may fail for polyphone words problem, such as “朝阳” can be read as “chao2yang2” and “zhao1yang2”, and Out of Vocabulary(OOV) problem. CRF and RNN perform well for polyphone pronunciation selection with context features. However, CRF needs manually designed context features. Neural network always has more parameters, larger model size and more expensive computation cost. In the application for embedded device, small model size and quick computation are necessary. So compression and acceleration of neural network is a hot research field in recent years. There are several methods to compress deep learning model: low precision of data format, quantifica- tion, pruning, low rank factorization, and knowledge distillation. Low precision of data format replaces double or float with 16-bit float, which can reduce model size by quarter or half, but it cannot reduce running time when corresponding computations are not supported by existing instruction sets. Quan- tification and pruning methods pack some weights into one and prune the weights close to zero. It can ©2020 China National Conference on Computational Linguistics Published under Creative Commons Attribution 4.0 International License CCL 2020 Proceedi ngs of t he 19t h Chi na Nat i onal Conf …

Content truncated to 500 words