GPU-Friendly Local Regression for Voice Conversion

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1334–1338, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics GPU-Friendly Local Regression for Voice Conversion Taylor Berg-Kirkpatrick Dan Klein Computer Science Division University of California, Berkeley {tberg,klein}@cs.berkeley.edu Abstract Voice conversion is the task of transforming a source speaker’s voice so that it sounds like a target speaker’s voice. We present a GPU- friendly local regression model for voice con- version that is capable of converting speech in real-time and achieves state-of-the-art ac- curacy on this task. Our model uses a new approximation for computing local regression coefficients that is explicitly designed to pre- serve memory locality. As a result, our infer- ence procedure is amenable to efficient imple- mentation on the GPU. Our approach is more than 10X faster than a highly optimized CPU- based implementation, and is able to convert speech 2.7X faster than real-time. 1 Introduction Voice conversion is the task of transforming an ut- terance from a source speaker’s voice into a target speaker’s voice. The primary setup in recent work has been to learn this transformation from a paral- lel corpus consisting of recordings of the same se- quence of sentences read by both source and tar- get speakers (Stylianou et al., 1998). The converted speech is evaluated by how well its spectral proper- ties match those of the target voice. While various models have been proposed (Stylianou et al., 1998; Toda et al., 2007; Toda et al., 2005), the most accurate ones are non-parametric because the mapping between two voices’ spectra can be highly non-linear (Helander et al., 2012; Popa et al., 2012). Unfortunately, while non-parametric methods are accurate, they are also slow – current non-parametric approaches to voice conversion are too compute-intensive for the real-time speed re- quired by many voice conversion applications. In this paper, we begin with the state-of-the-art local linear regression (LLR) model used by by Popa et al. (2012) for voice conversion, and present a new GPU-based inference approach that greatly acceler- ates it, to much faster than real-time. LLR, in principle, requires each new model pre- diction to be a function of the entire set of training examples. In practice, LLR depends most strongly on nearby points, so a standard CPU implementa- tion will skip distant points, with limited loss of ac- curacy. A GPU cannot exploit sparsity in the same way (scan and skip) without suffering from memory bottlenecks, but even a GPU will be relatively slow if all training points are included in each computation. Our primary algorithmic change is to make use of a new sparsity structure that allows the GPU to skip major sections of the training data while still using dense memory access patterns on the points it does process. In experiments, this inference technique is more than 10X faster than a highly-optimized CPU- based implementation, operates almost three times faster than real-time, and is only slightly less accu- rate than the CPU-based …

Content truncated to 500 words