Abstract: Devices and techniques are generally described for compression of machine learning models. In some examples, a first weight value wi of a machine learning model is determined. The first weight value wi may be associated with a first region comprising an interval of numbers. A first loss for the first weight value wi may be determined using a first loss function comprising a weighting term comprising a sinusoid function with an argument of ??rwi. ? may be selected such that a maximum value of the weighting term for the first region is of a format that is compatible with first machine learning accelerator. A gradient of the first loss function may be determined. A second weight value wi may be determined using the gradient and the first loss. The second weight value wi may be stored in a first memory.

Abstract: Disclosed is a speech processing apparatus and method using a densely connected hybrid neural network. The speech processing method includes inputting a time domain sample of N*1 dimension for an input speech into a densely connected hybrid network; passing the time domain sample through a plurality of dense blocks in a densely connected hybrid network; reshaping the time domain samples into M subframes by passing the time domain samples through the plurality of dense blocks; inputting the M subframes into gated recurrent unit (GRU) components of N/M-dimension; outputting clean speech from which noise is removed from the input speech by passing the M subframes through GRU components.

Abstract: A method and apparatus for processing an audio signal are disclosed. According to an example embodiment, a method of processing an audio signal may include acquiring a final audio signal for an initial audio signal using a plurality of neural network models generating output audio signals by encoding and decoding input audio signals, calculating a difference between the initial audio signal and the final audio signal in a time domain, converting the initial audio signal and the final audio signal into Mel-spectra, calculating a difference between the Mel-spectra of the initial audio signal and the final audio signal in a frequency domain, training the plurality of neural network models based on results calculated in the time domain and the frequency domain, and generating a new final audio signal distinguished from the final audio signal from the initial audio signal using the trained neural network models.

Abstract: Disclosed are a method for coding a residual signal of LPC coefficients based on collaborative quantization and a computing device for performing the method. The residual signal coding method includes: generating encoded LPC coefficients and LPC residual signals by performing LPC analysis and quantization on an input speech; Determining a predicted LPC residual signal by applying the LPC residual signal to cross module residual learning; Performing LPC synthesis using the coded LPC coefficients and the predicted LPC residual signal; It may include the step of determining an output speech that is a synthesized output according to a result of performing the LPC synthesis.

Abstract: Provided is a training method of a neural network that is applied to an audio signal encoding method using an audio signal encoding apparatus, the training method including generating a masking threshold of a first audio signal before training is performed, calculating a weight matrix to be applied to a frequency component of the first audio signal based on the masking threshold, generating a weighted error function obtained by correcting a preset error function using the weight matrix, and generating a second audio signal by applying a parameter learned using the weighted error function to the first audio signal.

Abstract: Disclosed are an audio signal encoding method and audio signal decoding method, and an encoder and decoder performing the same. The audio signal encoding method includes applying an audio signal to a training model including N autoencoders provided in a cascade structure, encoding an output result derived through the training model, and generating a bitstream with respect to the audio signal based on the encoded output result.

Abstract: Disclosed is a speech processing apparatus and method using a densely connected hybrid neural network. The speech processing method includes inputting a time domain sample of N*1 dimension for an input speech into a densely connected hybrid network; passing the time domain sample through a plurality of dense blocks in a densely connected hybrid network; reshaping the time domain samples into M subframes by passing the time domain samples through the plurality of dense blocks, inputting the M subframes into gated recurrent unit (GRU) components of N/M-dimension; outputting clean speech from which noise is removed from the input speech by passing the M subframes through GRU components.

Abstract: A method and apparatus for processing an audio signal are disclosed. According to an example embodiment, a method of processing an audio signal may include acquiring a final audio signal for an initial audio signal using a plurality of neural network models generating output audio signals by encoding and decoding input audio signals, calculating a difference between the initial audio signal and the final audio signal in a time domain, converting the initial audio signal and the final audio signal into Mel-spectra, calculating a difference between the Mel-spectra of the initial audio signal and the final audio signal in a frequency domain, training the plurality of neural network models based on results calculated in the time domain and the frequency domain, and generating a new final audio signal distinguished from the final audio signal from the initial audio signal using the trained neural network models.

Abstract: Disclosed are a method for coding a residual signal of LPC coefficients based on collaborative quantization and a computing device for performing the method. The residual signal coding method includes: generating encoded LPC coefficients and LPC residual signals by performing LPC analysis and quantization on an input speech; Determining a predicted LPC residual signal by applying the LPC residual signal to cross module residual learning; Performing LPC synthesis using the coded LPC coefficients and the predicted LPC residual signal; It may include the step of determining an output speech that is a synthesized output according to a result of performing the LPC synthesis.

Abstract: Disclosed are an audio signal encoding method and audio signal decoding method, and an encoder and decoder performing the same. The audio signal encoding method includes applying an audio signal to a training model including N autoencoders provided in a cascade structure, encoding an output result derived through the training model, and generating a bitstream with respect to the audio signal based on the encoded output result.

Abstract: Provided is a training method of a neural network that is applied to an audio signal encoding method using an audio signal encoding apparatus, the training method including generating a masking threshold of a first audio signal before training is performed, calculating a weight matrix to be applied to a frequency component of the first audio signal based on the masking threshold, generating a weighted error function obtained by correcting a preset error function using the weight matrix, and generating a second audio signal by applying a parameter learned using the weighted error function to the first audio signal.