IC data from Drakopoulos et al., 2026

This dataset contains the inferior colliculus (IC) recordings and simulations that were used to evaluate the ICNet and AidNet models in Drakopoulos et al., 2026. An example AidNet model is available via https://doi.org/10.5281/zenodo.18407090 or https://github.com/fotisdr/AidNet_example. A Jupyter notebook is included with a simple usage example for the AidNet model and sound examples. 

Drakopoulos, F., Pellatt, L., Sabesan, S., Xia, Y., Gong, T., Fragner, A., & Lesica, N. A. (2026). Optimal hearing aid design through restoration of the neural code. bioRxiv 2026.02.02.703273; doi: 10.64898/2026.02.02.703273

ICNet evaluation

The ICNet_sound_data and ICNet_MUA_data matfiles can be accessed on MATLAB or Python and contain the data corresponding to the 4 sounds that were used in ICNet model evaluation (Fig. 1c). Each sound segment is 30 s in duration and is calibrated with a 2e-5 Pa reference:

speech_in_quiet: A speech segment from the UCL SCRIBE dataset consisting of sentences spoken by a male talker and presented at 60 dB SPL.

speech_in_noise: A speech segment from the UCL SCRIBE dataset consisting of sentences spoken by a female talker and presented at 85 dB SPL. The speech segment was mixed in hallway noise from the Microsoft Scalable Noisy Speech dataset at 0 dB SNR.

moving_ripples: Dynamic moving ripples with frequency sinusoids between 4.7 kHz and 10.8 kHz presented at 85 dB SPL.

music: Three seconds from each of 10 mixed pop songs from the musdb18 dataset presented at 75 dB SPL.

fs: The sampling rate of the sounds (24414.0625 Hz).

The ICNet_MUA_data matfile contains the real and simulated IC activity elicited by each of the 4 sounds on two successive trials. The data is structured in 4 fields (true_trial1, true_trial2, model_trial1 and model_trial2) that contain the respective normal-hearing (NH) and hearing-impaired (HI) responses to the 4 sounds. Each response corresponds to a matrix of size [animals x time bins x channels] that contains multi-unit spikes across time bins and channels for each of the 10 NH and 10 HI animals. Each time element of the matrix indicates the spike count in one 1.3 ms bin (exactly 32 samples of the corresponding audio). The sampling rate of the recorded multi-unit activity is also provided (fs = 24414.0625 / 32 = 762.9395 Hz). 

AidNet in vivo evaluation

The AidNet_invivo_sound_data and AidNet_invivo_MUA_data matfiles can be accessed on MATLAB or Python and contain the data corresponding to the 4 sounds that were used in the AidNet in vivo evaluation (Fig. 4c and Supplementary Fig. 6b). The four sounds used in ICNet evaluation are provided before and after processing with each sound strategy and are structured as follows:

unprocessed: The four sounds speech_in_quiet, speech_in_noise, moving_ripples and music are provided without any processing (repeated from the matfile ICNet_sound_data). 

nalrp: The four sounds are provided for each animal of the in vivo evaluation after processing with the NAL-RP linear amplification strategy. Each data structure is a matrix of size [animals x time] that contains the processed sound for each of the 3 HI animals.

nalnl2: The four sounds are provided after processing with the NAL-NL2 compressive amplification strategy following the same format. 

aidnet: The four sounds are provided after processing with the respective AidNet model following the same format.

fs: The sampling rate of the sounds (24414.0625 Hz).

The AidNet_invivo_MUA_data matfile contains the real and simulated IC activity elicited by each of the unprocessed and processed sounds. The data is given for each of the sound processing strategies explained above and is structured in 2 fields (model and true) that contain the respective normal-hearing (NH) and hearing-impaired (HI) responses to each sound after averaging across repeated presentations and after MCA alignment. Each response corresponds to a matrix of size [animals x time bins x channels] that contains multi-unit spike means across time bins and channels for each of the 9 NH and 3 HI animals.

AidNet in silico evaluation

The AidNet_insilico_sound_data and AidNet_insilico_MUA_data matfiles can be accessed on MATLAB or Python and contain the data corresponding to the 4 sounds that were used in the AidNet in silico evaluation (Fig. 4d). The four sounds used in ICNet evaluation are provided before and after processing with each sound strategy and are structured as follows:

unprocessed: The four sounds speech_in_quiet, speech_in_noise, moving_ripples and music are provided without any processing. 

nalrp: The four sounds are provided for each animal of the in vivo evaluation after processing with the NAL-RP linear amplification strategy. Each data structure is a matrix of size [animals x time] that contains the processed sound for each of the 14 HI animals.

nalnl2: The four sounds are provided after processing with the NAL-NL2 compressive amplification strategy following the same format. 

aidnet: The four sounds are provided after processing with the respective AidNet model following the same format.

aidnetswap: The four sounds are provided after processing with the swapped AidNet model following the same format.

fs: The sampling rate of the sounds (24414.0625 Hz).

The AidNet_insilico_MUA_data matfile contains the simulated IC activity elicited for all NH and HI animals by each of the unprocessed and processed sounds. The model data is given for each of the sound processing strategies explained above and contains the expectation over counts for each respective NH and HI ICNet model after MCA alignment. Each response corresponds to a matrix of size [animals x time bins x channels] that contains multi-unit spike means across time bins and channels for each of the 9 NH and 14 HI animals. 

For questions, please reach out to one of the corresponding authors:

Fotios Drakopoulos: [e-mail]

Nicholas A Lesica: [e-mail]

This work was supported by UK MRC UKRI3206, EPSRC EP/W004275/1, BBSRC BB/Y008758/1 and MRC MR/W019787/1.