bfold =Wfold =・・・・・・γ+√ σ 2+εε·4γ+√ σ 2Reducing computational load by simplifying the formulaVisualizing the AI NR functionSuppressing howling in initial transfer function measurementsBatch Normalization Folding is a technique used to implement AI-powered functions in hearing aids, which have limited computational capacity and memory. By mathematically modifying and combining the calculations performed across two distinct layers in advance, we can completely eliminate the computational load and parameters required for one of the layers.The AI, deep neural network (DNN) analyzes the sounds entering the hearing aid, a mix of human voices and background noise. The AI NR instantly suppresses just the noise and delivers natural sound to the user’s ears. It excels at suppressing transient and impulsive noises.Input sound: Human voice + noiseDNN analyzes the input soundThe sound balance determined by AI NRRed: retained sound; blue: suppressed soundReducing thecomputational loady = Wfoldx + bfold(b – μ) + βproviding comfort while wearing hearing aids is preventing howling. This is a phenomenon that occurs when sound escap-ing through the gap between the hearing aid and the ear enters the microphone, causing repeated amplification and producing disruptive noise (a high-pitched whistling sound). This sound is strikingly unpleasant for the user and can lower their motivation to wear hearing aids. This is where Yuuki Yuno comes in. She was placed in charge of feedback cancellation to suppress this howling, during the development of the new hearing aid.“The difficult part of developing the feedback cancellation system was that it had to be capable of adapting to various sound environments,” explains Yuno. “The range of environments that can be verified in simulations is limited. Even if we used different types of noise as input, the actual feedback path inevitably varies, depending on the environment where the sound is heard and on factors like reflection. The process of repeating the verification process to check whether the implemented solution would be effective in each of the different cases required lots of patience.” The basic methods for feedback suppression involve sub-tracting an estimated feedback signal from the input signal or applying frequency shifting. These methods were applied in previous models of the Rionet series, but caused some deg-radation of sound quality and additional unwanted noise. So, with this model, Yuno focused on adjustments based on the initial transfer function measurement. Here, the initial transfer function refers to the frequency response of the path from the sound emitted by the hearing aid’s earphone to the microphone, measured while the user wears the hearing aid. Measuring the Pre-combining the calculations for the two layersz – μ+√ σ 2εβ+· W ,z = Wx + by = γ·Sato was, however, well aware of the significant hurdles to commercialization.“AI generally performs better at higher computational load. But hearing aids require low computational load and minimal memory usage. So we couldn’t simply implement high-perfor-mance AI as is. The DSP also posed constraints, related mainly to whether it could achieve the required computational efficien-cy. We had to consider all these factors to find the best possible solution.”There were two major challenges. The first was the problem of limited internal memory of hearing aids. The second was the constraint on computational power. To suppress noise, each calculation had to be completed in just 3 milliseconds (0.003 sec-onds). While the AI NR could demonstrate its full potential in computer-based simulations, lower computational speed of the DSP installed in a small hearing aid prevented it from producing equivalent performance.“We explored various possible solutions and focused on optimizing the efficiency of the mathematical calculations. Then we implemented them in DSP computations to reduce the com-putational load required for the same processing task. We made slow but steady progress and finally succeeded in enhancing the required performance and computational efficiency, even with the limitations in memory capacity.”In addition to reducing noise, another function crucial for The original formulaInputFully connected layerBatch normalization layerAfter modificationInputFully connected layerBatch normalization layer
元のページ ../index.html#6