Conceptual

RingFormer Neural Vocoder with Ring Attention and Conformer for Speech Synthesis

RingFormer is a neural vocoder that turns mel-spectrograms into speech waveforms by combining a convolution-augmented transformer (Conformer) with ring attention, which processes long sample-level sequences block by block to capture both local and global structure at low computational cost. Trained adversarially with dual discriminators and used as the decoder in the VITS text-to-speech pipeline, it delivers real-time waveform generation competitive with GAN vocoders such as HiFi-GAN and BigVGAN.