Ear-worn device with neural network for noise reduction and/or spatial focusing using multiple input audio signals
An ear-worn device may include two or more microphones configured to generate time-domain audio signals, each of the two or more microphones configured to generate one of the time-domain audio signals; processing circuitry comprising analog processing circuitry, digital processing circuitry, beamforming circuitry, and short-time Fourier transformation (STFT) circuitry, the processing circuitry configured to generate, from the time-domain audio signals, one or more frequency-domain non-beamformed audio signals and one or more frequency-domain beamformed signals; and enhancement circuitry comprising neural network circuitry configured to receive multiple frequency-domain input audio signals originating from the one or more frequency-domain non-beamformed audio signals and the one or more frequency-domain beamformed signals, and implement a single neural network trained to generate, based on the multiple frequency-domain input audio signals, a noise-reduced and spatially-focused output audio signal or an output for generating a noise-reduced and spatially-focused output audio signal.
1. An ear-worn device, comprising:
two or more microphones configured to generate time-domain audio signals, each of the two or more microphones configured to generate one of the time-domain audio signals;
processing circuitry comprising analog processing circuitry, digital processing circuitry, beamforming circuitry, and short-time Fourier transformation (STFT) circuitry, the processing circuitry configured to generate, from the time-domain audio signals, one or more frequency-domain non-beamformed audio signals and one or more frequency-domain beamformed signals, the one or more frequency-domain non-beamformed audio signals each comprising a speech portion and a noise portion, and the one or more frequency-domain beamformed audio signals each comprising a speech portion and a noise portion; and
neural network circuitry downstream of the beamforming circuitry, the neural network circuitry configured to:
receive multiple frequency-domain input audio signals originating from the one or more frequency-domain non-beamformed audio signals and the one or more frequency-domain beamformed signals; and
implement a single neural network trained to generate, based on the multiple frequency-domain input audio signals, a noise-reduced and spatially-focused output audio signal or an output configured to generate the noise-reduced and spatially-focused output audio signal.
2. The ear-worn device of claim 1 , wherein the one or more frequency-domain beamformed signals include a frequency-domain beamformed signal having a dipole, hypercardioid, supercardioid, or cardioid directional pattern.
3. The ear-worn device of claim 1 , wherein the two or more microphones comprise a front microphone and a back microphone, and the one of more frequency-domain non-beamformed audio signals comprise a frequency-domain non-beamformed audio signal originating from the front microphone and a frequency-domain non-beamformed audio signal originating from the back microphone.
4. The ear-worn device of claim 1 , wherein the single neural network comprises a recurrent network.
5. The ear-worn device of claim 1 , further comprising interpolation circuitry configured to interpolate between:
one of the multiple frequency-domain input audio signals; and
the noise-reduced and spatially-focused output audio signal, or a processed version thereof.
6. The ear-worn device of claim 1 , wherein the noise-reduced and spatially-focused output audio signal uses a mapping of gains to respective spatial regions.
7. The ear-worn device of claim 6 , wherein the mapping is predetermined.
8. The ear-worn device of claim 7 , wherein the mapping of the gains to the respective spatial regions comprises applying a gain of 1 to audio generated from sounds coming from a target spatial region and applying a gain of 0 to audio generated from sounds coming from other spatial regions.
9. The ear-worn device of claim 8 , wherein the target spatial region has an angle relative to a wearer of the ear-worn device of approximately equal to or between 10-180 degrees.
10. The ear-worn device of claim 7 , wherein the mapping of the gains to the respective spatial regions comprises mapping more than two spatial regions each to a different gain, and one or more of the spatial regions are processed with gains not equal to 1 or 0.
11. The ear-worn device of claim 6 , wherein the mapping is not predetermined.
12. The ear-worn device of claim 11 , wherein the output for generating the noise-reduced and spatially-focused output audio signal comprises a sound map indicating frequency components originating from each of multiple spatial regions.
13. The ear-worn device of claim 12 , wherein the ear-worn device further comprises enhancement circuitry that comprises the neural network circuitry, and the enhancement circuitry is further configured to apply a beam pattern to the sound map, and the beam pattern is based on a selection from a wearer of the ear-worn device.
14. The ear-worn device of claim 13 , wherein the selection from the wearer of the ear-worn device comprises a selection of a size of a front-facing spatial region to use for focusing.
15. The ear-worn device of claim 11 , wherein the ear-worn device further comprises enhancement circuitry that comprises the neural network circuitry, the output for generating the noise-reduced and spatially-focused output audio signal comprises values calculated for a metric from audio from multiple beams, each of the multiple beams pointing at a different angle around a wearer of the ear-worn device, and the enhancement circuitry is configured to combine the audio from the multiple beams using the values for the metric.
16. The ear-worn device of claim 1 , wherein the neural network is trained on both captured data and synthetic data.
17. The ear-worn device of claim 1 , further comprising:
an inertial measurement unit (IMU); and
second processing circuitry configured to track head movements of a wearer of the ear-worn device using measurements from the IMU and cause an absolute coordinate system to be used for the spatial focusing based on the head movements.
18. The ear-worn device of claim 1 , wherein:
the ear-worn device further comprises enhancement circuitry that comprises the neural network circuitry;
the processing circuitry is coupled between the two or more microphones and the enhancement circuitry;
the analog processing circuitry is coupled between the two or more microphones and the digital processing circuitry;
the digital processing circuitry is coupled between the analog processing circuitry and the beamforming circuitry;
the beamforming circuitry is coupled between the digital processing circuitry and the STFT circuitry;
the analog processing circuitry is configured to perform one or more of analog preamplification, analog filtering, and analog-to-digital conversion; and
the digital processing circuitry is configured to perform one or more of wind reduction, input calibration, and anti-feedback processing.
19. The ear-worn device of claim 1 , wherein:
the ear-worn device further comprises enhancement circuitry that comprises the neural network circuitry;
the processing circuitry is coupled between the two or more microphones and the enhancement circuitry;
the analog processing circuitry is coupled between the two or more microphones and the digital processing circuitry;
the digital processing circuitry is coupled between the analog processing circuitry and the STFT circuitry;
the STFT circuitry is coupled between the digital processing circuitry and the beamforming circuitry;
the analog processing circuitry is configured to perform one or more of analog preamplification, analog filtering, and analog-to-digital conversion; and
the digital processing circuitry is configured to perform one or more of wind reduction, input calibration, and anti-feedback processing.
20. The ear-worn device of claim 1 , wherein the neural network circuitry is configured to output a single output based on the multiple frequency-domain input audio signals.