Acoustic models based on long short-term memory recurrent neural networks (LSTM-RNNs) were applied to statistical parametric speech synthesis (SPSS) and showed significant improvements in naturalness and latency over those based on hidden Markov models (HMMs). This paper describes further optimizations of LSTM-RNN-based SPSS for deployment on mobile devices; weight quantization, multi-frame inference, and robust inference using an ε-contaminated Gaussian loss function. Experimental results in subjective listening tests show that these optimizations can make LSTM-RNN-based SPSS comparable to HMM-based SPSS in runtime speed while maintaining naturalness. Evaluations between LSTM-RNN- based SPSS and HMM-driven unit selection speech synthesis are also presented.

Some modern superscalar microprocessors provide only imprecise exceptions. That is, they do not guarantee to report the same exception that would be encountered by a straightforward sequential execution of the program. In exchange, they offer increased performance or decreased chip area (which amount to much the same thing).This performance/precision tradeoff has not so far been much explored at the programming language level. In this paper we propose a design for imprecise exceptions in the lazy functional programming language Haskell. We discuss several designs, and conclude that imprecision is essential if the language is still to enjoy its current rich algebra of transformations. We sketch a precise semantics for the language extended with exceptions.The paper shows how to extend Haskell with exceptions without crippling the language or its compilers. We do not yet have enough experience of using the new mechanism to know whether it strikes an appropriate balance between expressiveness and performance.