README.md

metadata

license: mit
task_categories:
  - audio-to-audio
language:
  - en

Dataset Card for LenslessMic Version of Librispeech Dataset

Dataset Summary

A LenslessMic version of the Librispeech dataset from the "LenslessMic: Audio Encryption and Authentication via Lensless Computational Imaging" paper.

Partition	# Audio	# Frames
train-clean	587	73,699
train-other	150	18,561
test-clean	1,089	185,773
test-other	512	62,901

To download the dataset and work with it, use our official repository.

Dataset is collected using DigiCam. Setup configuration:

Parameter	Value
Screen Size	[1920, 1200]
Screen Pixel-Pitch	0.27 mm
Screen-To-Mask Distance	30e-2 m
Sensor Size	[4056, 3040]
Sensor Size Downsample Coefficient	8
Sensor Pixel-Pitch	1.55 × 10⁻⁶ m
Mask-To-Sensor Distance	≈ 4e-3 m
Image size on the Screen (256 case)	928 × 928
Image size on the Screen (288 case)	1044 × 1044
Vertical Shift on the Screen (256 case)	-23
Vertical Shift on the Screen (288 case)	-20
Number of masks	100
Mask Aperture Shape (for 1/3 channels)	[18, 24]
Mask Center	[55, 77]

For other configuration, please refer to the codebase above.

Dataset Structure

Dataset is structured in the following format:

.
└── partition_name
    ├── audio # for original audio
    |   ├── filename_i.flac # i-th audio file
    |   └── filename_i.txt  # transcription of the i-th audio file
    └── neural_audio_codec_name
        ├── lensed # lensed version of the video representation
        |   ├── filename_i.mkv # normalized video representation of i-th audio file using this codec
        |   ├── filename_i_max_vals.pth # per-frame max values before normalization
        |   └── filename_i_min_vals.pth # per-frame min values before normalization
        └── lensless_measurement # lensless version captured using LenslessMic
            ├── filename_i.mkv # lensless video of the i-th audio file
            ├── filename_i.txt # label 'j' of the mask from the masks dir used for this video
            └── masks # masks for the lensless camera
                └── mask_j.npy # mask pattern

Codecs used:

1. Original 16khz DAC: 32x32_120_16khz_original.
2. Fine-tuned DAC with additional 13-th codebook and 256 latent size instead of 1024: 16x16_130_16khz.
3. X-codec: 32x32_80_16khz_xcodec_hubert_general_audio.

Some codecs have different types of lensless measurements:

1. lensless_measurement: standard version. Resizes images in a screen in a such a way that they have size 256x256 on the sensor.
2. lensless_measurement_group_n_m_r_c: groupped version, i.e., frames are groupped into n*m grid with additional r pixels between rows and c pixels between columns. Still 256x256 on the sensor.
3. lensless_measurement288x288_group_n_m_r_c: the same as standard groupped version but 288x288 image on the sensor.

Region of interest for the reconstruction for this dataset is:

Sensor Image Size	Top Left Corner	Height	Width
256 x 256	[65, 118]	256	256
288 x 288	[47, 103]	288	288

Citation

If you use this dataset, 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},
}