Constant improvement of hearing aids by retraining with selected data collected in use
AssigneeFortell Research Inc.
InventorsIgor Lovchinsky, Jonathan Macoskey, Philip Meyers, IV, Israel Malkin, Nicholas Morris
Life of this patent
Estimating the term…
Abstract
Ear-worn devices such as hearing aids are described herein. The ear-worn devices include a machine learning model for enhancing received audio signals to be played out to the wearer. The machine learning model is re-trainable based on the received audio signals. The ear-worn device is configured to selectively store received audio signals to be used in re-training the machine learning model.
Claims28
A hearing aid having a de-noising model trainable with data collected by the hearing aid during operation, the hearing aid comprising:
- a microphone configured to receive an acoustic signal during operation;
- random access memory configured to temporarily store the acoustic signal;
- processing circuitry configured to determine a metric of the acoustic signal and to de-noise the acoustic signal using a neural network;
- and
- persistent memory, wherein the processing circuitry is additionally configured to selectively send the acoustic signal to the persistent memory depending on the metric.
An ear-worn device, comprising:
- a microphone configured to receive an acoustic signal;
- processing circuitry configured to make a determination of whether to store or discard data obtained from the acoustic signal based on a metric obtained using the data;
- and
- a storage configured to store the data based on the determination being to store the data, wherein the data is used to train a machine learning model implemented by the processing circuitry.
A method of manufacturing a hearing aid having a de-noising model trainable with data collected by the hearing aid during operation, the method comprising:
- arranging a microphone to receive an acoustic signal during operation;
- arranging random access memory to temporarily store the acoustic signal;
- configuring processing circuitry to determine a metric of the acoustic signal and to de-noise the acoustic signal using a neural network;
- coupling persistent memory to the processing circuitry;
- and
- configuring the processing circuitry to selectively send the acoustic signal to the persistent memory depending on the metric.
A method of collecting training data, with an ear-worn device having a microphone, the method comprising:
- receiving an acoustic signal with the microphone of the ear-worn device;
- after receiving the acoustic signal, writing a digital representation of the acoustic signal to a buffer;
- calculating a metric of the acoustic signal based on the digital representation of the acoustic signal;
- comparing the metric to a threshold value;
- writing the digital representation of the acoustic signal to persistent storage in dependence on a result of comparing the metric to the threshold value;
- and
- subsequently transferring the digital representation of the acoustic signal from the ear-worn device to an external device.
Dependencies are inferred from the claim text.