Fixed Point Analysis Workflow for efficient Design of Convolutional Neural Networks in Hearing Aids

Neural networks (NN) are a powerful tool to tackle complex problems in hearing aid research, but their use on hearing aid hardware is currently limited by memory and processing power. To enable the training with these constrains, a fixed point analysis and a memory friendly power of two quantization (replacing multiplications with shift operations) scheme has been implemented extending TensorFlow, a standard framework for training neural networks, and the Qkeras package [1, 2]. The implemented fixed point analysis detects quantization issues like overflows, underflows, precision problems and zero gradients. The analysis is done for each layer in every epoch for weights, biases and activations respectively. With this information the quantization can be optimized, e.g. by modifying the bit width, number of integer bits or the quantization scheme to a power of two quantization. To demonstrate the applicability of this method a case study has been conducted. Therefore a CNN has been trained to predict the Ideal Ratio Mask (IRM) for noise reduction in audio signals. The dataset consists of speech samples from the TIMIT dataset mixed with noise from the Urban Sound 8kand VAD-dataset at 0 dB SNR. The CNN was trained in floating point, fixed point and a power of two quantization. The CNN architecture consists of six convolutional layers followed by three dense layers. From initially 1.9 MB memory footprint for 468k float32 weights, the power of two quantized network is reduced to 236 kB, while the Short Term Objective Intelligibility (STOI) Improvement drops only from 0.074 to 0.067. Despite the quantization only a minimal drop in performance was observed, while saving up to 87.5 % of memory, thus being suited for employment in a hearing aid