Generative Modeling for Low Dimensional Speech Attributes with Neural Spline Flows
This work addresses a specific bottleneck in speech synthesis for applications requiring precise pitch control, representing an incremental improvement over existing methods.
The paper tackled the problem of fine-grained adjustability of low-dimensional, discontinuous pitch attributes in generative text-to-speech synthesis, achieving improved modeling by exploring techniques in Normalizing Flow models, with Neural Spline flows identified as a highly expressive alternative.
Despite recent advances in generative modeling for text-to-speech synthesis, these models do not yet have the same fine-grained adjustability of pitch-conditioned deterministic models such as FastPitch and FastSpeech2. Pitch information is not only low-dimensional, but also discontinuous, making it particularly difficult to model in a generative setting. Our work explores several techniques for handling the aforementioned issues in the context of Normalizing Flow models. We also find this problem to be very well suited for Neural Spline flows, which is a highly expressive alternative to the more common affine-coupling mechanism in Normalizing Flows.