license: apache-2.0

SphereVAE

SphereVAE: Hyperspherical Latent Autoencoders for Robust Autoregressive Speech Representation Modeling

Model

SphereVAE combines a SEANet audio codec with causal Transformers and a Power-spherical latent space for robust speech representation modeling.

Installation

Use Python 3.10 or newer. Install a PyTorch and torchaudio pair appropriate for your machine, then install the dependencies:

python -m pip install -r requirements.txt

GPU training uses Accelerate. Configure it for your machine before launching:

accelerate config

Training data

Place a Kaldi-style SCP file at data/train.scp, or update data.train_scp_path in configs/config_Sphere_VAE.yaml:

utt_0001 /path/to/audio_0001.flac
utt_0002 /path/to/audio_0002.wav

Each line contains an utterance ID and an audio path. Audio is processed as mono 24 kHz input and prepared as training segments by the dataset pipeline.

Training

CUDA_VISIBLE_DEVICES=0 accelerate launch train.py -c configs/config_Sphere_VAE.yaml

Adjust training settings in configs/config_Sphere_VAE.yaml. Checkpoints and TensorBoard logs are written to the configured experiment directory. For multi-GPU training, configure Accelerate and set the visible devices as needed.

Inference

python infer.py \
  --checkpoint exps/Sphere_VAE/G_400000.pth \
  --input data/test \
  --output output/reconstructed \
  --device cuda:0

--input accepts an audio file or directory. Supported extensions are WAV, FLAC, MP3, OGG, and M4A. Use --config to select another configuration. Inference uses posterior sampling, so repeated runs can produce different reconstructions. Output is trimmed to the input length.

Layout

models/model_Sphere_VAE.py  Sphere_VAE model and builder
modules/                SEANet, Transformer, spherical distribution, utilities
train.py                  Training entrypoint
infer.py                  Audio reconstruction entrypoint
dataset.py              SCP-listed audio dataset
configs/                Sphere_VAE configuration
discriminators/         STFT discriminator
losses/                 Spectral and adversarial losses
utils/                   Configuration, checkpoints, and compilation utilities

Citation

@misc{zhang2026spherevaehypersphericallatentautoencoders,
  title={SphereVAE: Hyperspherical Latent Autoencoders for Robust Autoregressive Speech Representation Modeling},
  author={Haoyu Zhang and Jingbin Hu and Hanke Xie and Qirui Zhan and Wenhao Li and Ziyu Zhang and Xiaming Ren and Yue Li and Xunyu Zhu and Zhipeng Chen and Lei Xie},
  year={2026},
  eprint={2609.09903},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  url={https://arxiv.org/abs/2609.09903}
}

Acknowledgments

Parts of the code are adapted from Kyutai Mimi and Meta AudioCraft/EnCodec, as indicated by the original notices retained in the source files. This export does not assign a new license to that third-party code. The source checkout did not include the root license files referenced by those notices; applicable upstream license texts still need to be supplied before release.

ASLP-lab/SphereVAE