This repository contains a deep-neural-network hearing-aid (DNN-HA) processing strategy that can provide individualised sound processing for the audiogram of a listener using a single DNN model architecture. The supporting paper can be found here and can be cited as follows:
F. Drakopoulos, A. Van Den Broucke and S. Verhulst, "A DNN-Based Hearing-Aid Strategy For Real-Time Processing: One Size Fits All," ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Rhodes Island, Greece, 2023, pp. 1-5, doi: 10.1109/ICASSP49357.2023.10094887.
The DNN-HA strategy uses an audio signal and an audiogram as inputs to process sound in real-time, such as to provide optimal HA processing that compensates for the elevated hearing thresholds of an individual listener. The CNN-HA-12layers folder contains the trained DNN-HA model, while an example script test_DNN-HA_wavfile.py is provided for running and evaluating the model. This repository also contains this README.md document, a license file, and the supplementary files that are necessary for the execution of the model.
The DNN-HA model was trained using the differentiable closed-loop framework that we describe here. The framework is based on CoNNear, a DNN version of a biophysically realistic model of the auditory system. The code for the CoNNear periphery model can be found here. More details about the DNN-HA model architecture and the training procedure will be made available upon publication of the corresponding paper as part of ICASSP 2023.
The test_DNN-HA_wavfile.py script provides a usage example of the DNN-HA model and can be used to process an example sentence (00131.wav) from the Flemish Matrix presented at 70 dB SPL. Within this script, the audiogram of a listener (L44) can be specified across 8 frequencies (125-8000 Hz) and be used to process a wavfile (L45) presented at the desired intensity level (L47). The script visualises the audio waveform before and after processing, and also the unprocessed and processed spectra using PyOctaveBand. By default, the processed sound is also saved as a wavfile under the wavfiles folder. The script also provides the choice to add noise at a desired SNR (L48) and to process sound in frames (L53), such that the DNN-HA model can be evaluated for real-time and/or low-latency processing. More information can be found in the test_DNN-HA_wavfile.py script.
To run the example script and the DNN-HA model in Python, Numpy, Scipy and Tensorflow are necessary. We used a conda environment (v4.14.0) that included the following versions:
- Python 3.9.13
- Numpy 1.21.5
- Scipy 1.7.3
- Tensorflow 2.7.0
If you use this code, please cite the corresponding paper or this repository:
Fotios Drakopoulos, Arthur Van Den Broucke, & Sarah Verhulst. (2023). DNN-HA: A DNN-based hearing-aid strategy for real-time processing (v1.0). Zenodo. https://doi.org/10.5281/zenodo.7717218
For questions, please reach out to one of the corresponding authors:
- Fotios Drakopoulos: f.drakopoulos@ucl.ac.uk
- Sarah Verhulst: s.verhulst@ugent.be
This work was supported by European Research Council ERC-StG-678120 (RobSpear) and FWO grant G063821N Machine Hearing 2.0.