Abstract: A method and an electronic device for training a complex neural model of acoustic echo cancellation (AEC) are provided. The method includes: generating an anchor audio pair, a positive audio pair and a negative audio pair according to multiple near-end signals and multiple acoustic echo signals; utilizing the complex neural model to extract an anchor audio feature, a positive audio feature and a negative audio feature from the anchor audio pair, the positive audio pair and the negative pair, respectively; calculating a loss function according to the anchor audio feature, the positive audio feature and the negative audio feature; and tuning at least one parameter of the complex neural model according to the loss function.

Abstract: A device operates to perform acoustic echo cancellation. The device includes a speaker to output a far-end signal at the device, a microphone to receive at least a near-end signal and the far-end signal from the speaker to produce a microphone output, and an AI accelerator operative to perform neural network operations according to a first neural network model and a second neural network model to output an echo-suppressed signal. The device further includes a digital signal processing (DSP) unit. The DSP unit is operative to perform adaptive filtering to remove at least a portion of the far-end signal from the microphone output to generate a filtered near-end signal, and perform Fast Fourier Transform (FFT) and inverse FFT (IFFT) to generate input to the first neural network model and the second neural network model, respectively.

Abstract: An acoustic echo cancellation (AEC) system includes an adaptive filter, a subtraction circuit, and a processor executing a model. The adaptive filter is arranged to generate an estimated echo signal according to a first microphone signal played by a loudspeaker. The subtraction circuit is arranged to subtract the estimated echo signal from a signal that is output from a microphone receiving both of a speech signal and an echo signal, to generate a second microphone signal, wherein the first microphone signal is not output from the microphone, and the echo signal is transmitted from the loudspeaker to the microphone. The model is arranged to perform short-time Fourier transform upon the first microphone signal and the second microphone signal, respectively, and generate an estimated speech signal through a neural network according to a first transformed microphone signal and a second transformed microphone signal.

Abstract: The acoustic echo cancellation (AEC) system includes a loudspeaker interface coupled to a loudspeaker, a microphone interface coupled to a microphone, and a processor executing a model. The model predicts and generates a spectral magnitude mask (SMM) through a neural network according to a first microphone signal output by the loudspeaker and a second microphone signal output by the microphone, wherein a noisy speech signal is a sum of a clean speech signal and a noise signal; the SMM is a ratio of a spectral magnitude of an estimated speech signal and a spectral magnitude of the noisy speech signal; a true mask is a ratio of a spectral magnitude of the clean speech signal and the spectral magnitude of the noisy speech signal; and the model applies a power function to a loss function of the model according to the true mask and a magnitude of the true mask.

Abstract: An acoustic echo cancellation (AEC) system includes a loudspeaker interface coupled to a loudspeaker, a microphone interface coupled to a microphone, and a processor executing a model. The model is arranged to predict and generate a spectral magnitude mask (SMM) through a neural network according to a first microphone signal output by the loudspeaker and a second microphone signal output by the microphone, wherein a noisy speech signal is a sum of a clean speech signal and a noise signal; the SMM is a ratio of a spectral magnitude of an estimated speech signal and a spectral magnitude of the noisy speech signal; a true mask is a ratio of a spectral magnitude of the clean speech signal and the spectral magnitude of the noisy speech signal; and a loss function of the model is a mean square error between the SMM and the true mask.

Abstract: A bit-widths determination method selects bit-widths for mixed-precision neural network computing on a target hardware platform. An activation quantization sensitivity (AQS) value is calculated for each convolution layer in a neural network. The AQS value indicates the sensitivity of convolution output to quantized convolution input. One or more convolution layers are grouped into a quantization group, which is to be executed by a corresponding set of target hardware. A group AQS value is calculated for each quantization group based on the AQS values of the convolution layers in the quantization group. Then bit-widths supported by the target hardware platform are selected for the corresponding quantization groups. The bit-widths are selected to optimize, under a given constraint, a sensitivity metric that is calculated based on each quantization group's group AQS value.

Abstract: A device operates to perform acoustic echo cancellation. The device includes a speaker to output a far-end signal at the device, a microphone to receive at least a near-end signal and the far-end signal from the speaker to produce a microphone output, and an AI accelerator operative to perform neural network operations according to a first neural network model and a second neural network model to output an echo-suppressed signal. The device further includes a digital signal processing (DSP) unit. The DSP unit is operative to perform adaptive filtering to remove at least a portion of the far-end signal from the microphone output to generate a filtered near-end signal, and perform Fast Fourier Transform (FFT) and inverse FFT (IFFT) to generate input to the first neural network model and the second neural network model, respectively.

Abstract: A compensator compensates for the distortions of a power amplifier circuit. A power amplifier neural network (PAN) is trained to model the power amplifier circuit using pre-determined input and output signal pairs that characterize the power amplifier circuit. Then a compensator is trained to pre-distort a signal received by the PAN. The compensator uses a neural network trained to optimize a loss between a compensator input and a PAN output, and the loss is calculated according to a multi-objective loss function that includes one or more time-domain loss function and one or more frequency-domain loss functions. The trained compensator performs signal compensation to thereby output a pre-distorted signal to the power amplifier circuit.

Abstract: An analog circuit is calibrated to perform neural network computing. Calibration input is provided to a pre-trained neural network that includes at least a given layer having pre-trained weights stored in the analog circuit. The analog circuit performs tensor operations of the given layer using the pre-trained weights. Statistics of calibration output from the analog circuit is calculated. Normalization operations to be performed during neural network inference are determined. The normalization operations incorporate the statistics of the calibration output and are performed at a normalization layer that follows the given layer. A configuration of the normalization operations is written into memory while the pre-trained weights stay unchanged.