CBHG

CBHG is a neural network module used in sequence modeling, particularly in neural speech synthesis. The name derives from its three main components: a Convolutional Bank, Highway networks, and a Gated Recurrent Unit (often bidirectional). The module is designed to convert input sequences such as characters, phonemes, or acoustic features into rich, high-level representations suitable for downstream decoding.

The Convolutional Bank consists of a set of one-dimensional convolutions with multiple kernel widths applied to the input. By stacking outputs from different kernel sizes, the bank captures local patterns at multiple time scales. This is typically followed by a pooling operation and a projection stage that reduces dimensionality. The Highway networks comprise several highway layers, which enable training of deep networks by allowing information to flow through transform and carry paths. Finally, the Bidirectional GRU (or another recurrent unit) models temporal dependencies across the sequence, producing a temporally aware, fixed-dimensional representation for each time step.

CBHG modules are used as feature extractors in Tacotron and related text-to-speech architectures. In Tacotron, a CBHG-based encoder processes input sequences into a sequence of high-level features that guide the attention-driven decoder. A variant known as the postnet CBHG is also employed to refine generated acoustic features, improving the quality of the final spectrogram outputs.

Overall, the CBHG design combines multi-scale local feature extraction with gated information flow and temporal modeling to produce robust representations for sequential data in speech synthesis and similar tasks.