README.md

metadata

license: mit
task_categories:
  - audio-to-audio
language:
  - en
datasets:
  - Blinorot/lensless_mic_librispeech
  - Blinorot/lensless_mic_random
  - Blinorot/lensless_mic_songdescriber

Model Card for LenslessMic Reconstruction Algorithms

Models Summary

Reconstruction algoritms from the "LenslessMic: Audio Encryption and Authentication via Lensless Computational Imaging" paper.

To download the models and work with them, use our official repository.

Models Details

The models are saved in the following format:

.
└── checkpoint_tag
    ├── checkpoint_name.pth # PyTorch checkpoint with model state dict under 'state_dict' key.
    └── config.yaml # Hydra config used to train the model

Checkpoint tag is represented in the following format:

{latent_size}_{training_dataset}_{loss_functions_used}_{reconstruction_algorithm}

1. The latent_size is either 16x16 or 32x32, depends on the neural audio codec used in the dataset.
2. The training dataset is either random or librispeech. For librispeech, a groupped version can be used, tagged as group_n_m_r_c (see LenslessMic Version of Librispeech (with 288x288 after group if the sensor image size is not the default 256x256). The version of the model, which is fine-tuned using train-other, is tagged as librispeech_other and _ft at the end.
3. The loss_function is usually MSE, SSIM, and Raw SSIM, as in the paper. We also provide checkpoints with only MSE, MSE and SSIM, and all three with L1 waveform or Mel Losses.
4. The reconstruction algorithm: PSF_Unet4M_U5_Unet4M is the Learned and R-Learned methods from the paper. Unet8M is the NoPSF method.

Citation

If you use these models, please cite it as follows:

@article{grinberg2025lenslessmic,
  title = {LenslessMic: Audio Encryption and Authentication via Lensless Computational Imaging},
  author = {Grinberg, Petr and Bezzam, Eric and Prandoni, Paolo and Vetterli, Martin},
  journal = {arXiv preprint arXiv:2509.16418},
  year = {2025},
}