Vocos

Instead of generating waveform samples directly in the time domain, Vocos predicts Fourier spectral coefficients and reconstructs audio through an inverse short-time Fourier transform, closing the quality gap between time-domain GAN vocoders and much faster frequency-domain synthesis. Single-pass generation makes it quick even on CPU. Two pretrained models on Hugging Face cover the main use cases: vocos-mel-24khz, a 13.5M parameter model trained for 1 million iterations on LibriTTS that decodes mel spectrograms, and vocos-encodec-24khz, a 7.9M parameter model trained on the DNS Challenge corpus that reconstructs audio from EnCodec tokens at multiple bandwidths. Installation is pip install vocos, with a train extra for reproducing models, and the library slots into Bark and other token-based TTS systems as a drop-in decoder. Gemelo AI released both code and weights under the MIT license, so commercial use is unrestricted, and the 2023 arXiv paper documents the architecture. The vocoder now appears as the decoder inside a range of open TTS and neural codec pipelines where inference speed matters as much as fidelity.