Google Inc.

Huawei Technologies Co., Limited

Koninklijke Philips Corporation N.v.

Ntt Corporation

Victoria University of Wellington

Telefonaktiebolaget Lm Ericsson (publ)

Gn Hearing A/s

Gn Resound A/s

Willem Bastiaan Kleijn

Jan K Skoglund

Minyue Li

Sze Chie Lim

Richard Heusdens

Janusz Klejsa

Turaj Zakizadeh Shabestary

Alejandro Luebs

Guoqiang Zhang

Wenyu Jin

Ermin Kozica

Richard C Hendriks

Dave Zachariah

Yue Lang

David Virette

Kenta Niwa

Christian Feldbauer

Nikolay D Gaubitch

Thomas Sherson

Volodya Grancharov

Stefan Bruhn

Wei Xiao

Alexander Ypma

David Yuheng Zhao

Peter Grosche

Steven Leonardus Josephus Dimphina Elisabeth Van De Par

Yiteng Huang

Andrew Allen

Andrew John MacDonald

Renat Vafin

Björn Volcker

Bjorn Volcker

Michael Takezo Chinen

Yuan Zeng

Bjoern Volcker

Bert Devries

Petko N Petkov

Mona Hakami

Andrew Storus

Daan Ratering

Chrisitian Feldbauer

Christos Tzagkarakis

Björn Völcker

Machine learning techniques which allow machine learning to be performed even when a cost function is not a convex function are provided. A machine learning system includes a plurality of node portions which learn mapping that uses one common primal variable by machine learning based on their respective input data while sending and receiving information to and from each other. The machine learning is performed so as to minimize, instead of a cost function of a non-convex function originally corresponding to the machine learning, a proxy convex function serving as an upper bound on the cost function. The proxy convex function is represented by a formula of a first-order gradient of the cost function with respect to the primal variable or by a formula of a first-order gradient and a formula of a second-order gradient of the cost function with respect to the primal variable.

Machine learning system, machine learning method, and program

Techniques for achieving a scalable DNN and a multi-task DNN, for example, are provided. A neural network system is a neural network system including a plurality of models. Each one of the plurality of models is a DNN including a plurality of layers. Some or all of the plurality of models include at least one layer for which some or all of model variables are equivalent or common (hereinafter referred to as a 'shared layer') and also include at least one layer for which the model variables are not equivalent nor common (hereinafter referred to as an “unshared layer”).

Scalable partially shared multi-task neural network system, method, and program

Implementations identify a small set of independent, salient features from an input signal. The salient features may be used for conditioning a generative network, making the generative network robust to noise. The salient features may facilitate compression and data transmission. An example method includes receiving an input signal and extracting salient features for the input signal by providing the input signal to an encoder trained to extract salient features. The salient features may be independent and have a sparse distribution. The encoder may be configured to generate almost identical features from two input signals a system designer deems equivalent. The method also includes conditioning a generative network using the salient features. In some implementations, the method may also include extracting a plurality of time sequences from the input signal and extracting the salient features for each time sequence.

Identifying salient features for generative networks

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for coding speech using neural networks. One of the methods includes obtaining a bitstream of parametric coder parameters characterizing spoken speech; generating, from the parametric coder parameters, a conditioning sequence; generating a reconstruction of the spoken speech that includes a respective speech sample at each of a plurality of decoder time steps, comprising, at each decoder time step: processing a current reconstruction sequence using an auto-regressive generative neural network, wherein the auto-regressive generative neural network is configured to process the current reconstruction to compute a score distribution over possible speech sample values, and wherein the processing comprises conditioning the auto-regressive generative neural network on at least a portion of the conditioning sequence; and sampling a speech sample from the possible speech sample values.

Speech coding using auto-regressive generative neural networks

A method includes receiving sampled audio data corresponding to utterances and training a machine learning (ML) model, using the sampled audio data, to generate a high-fidelity audio stream from a low bitrate input bitstream. The training of the ML model includes de-emphasizing the influence of low-probability distortion events in the sampled audio data on the trained ML model, where the de-emphasizing of the distortion events is achieved by the inclusion of a term in an objective function of the ML model, which term encourages low-variance predictive distributions of a next sample in the sampled audio data, based on previous samples of the audio data.

Trained generative model speech coding

An apparatus for providing active noise control, includes: one or more microphones configured to detect sound entering through an aperture of a building structure; a set of speakers configured to provide sound output for cancelling or reducing at least some of the sound; and a processing unit communicatively coupled to the set of speakers, wherein the processing unit is configured to provide control signals to operate the speakers, wherein the control signals are independent of an error-microphone output.

Wave-domain approach for cancelling noise entering an aperture

According to an aspect, a method for multi-channel echo cancellation includes receiving a microphone signal and a multi-channel loudspeaker driving signal. The multi-channel loudspeaker driving signal includes a first driving signal that drives a first loudspeaker, and a second driving signal that drives a second loudspeaker. The first driving signal is substantially the same as second driving signal. The microphone signal includes a near-end signal with echo. The method includes determining a unique solution for acoustic transfer functions for a present acoustic scenario based on the microphone signal and the multi-channel loudspeaker driving signal. The acoustic transfer functions include first and second acoustic transfer function. The unique solution is determined based on time-frequency transforms of observations from the present acoustic scenario and at least one previous acoustic scenario. The method includes removing the echo from the microphone signal based on the first and second acoustic transfer function.

Multi-channel echo cancellation with scenario memory

Provided are methods, systems, and apparatus for hierarchical decorrelation of multichannel audio. A hierarchical decorrelation algorithm is designed to adapt to possibly changing characteristics of an input signal, and also preserves the energy of the original signal. The algorithm is invertible in that the original signal can be retrieved if needed. Furthermore, the proposed algorithm decomposes the decorrelation process into multiple low-complexity steps. The contribution of these steps is generally in a decreasing order, and thus the complexity of the algorithm can be scaled.

Hierarchical decorrelation of multichannel audio

Techniques of source localization and acquisition involve a wideband joint acoustic source localization and acquisition approach in light of sparse optimization framework based on an orthogonal matching pursuit-based grid-shift procedure. Along these lines, a specific grid structure is constructed with the same number of grid points as compared to the on-grid case, but which is 'shifted' across the acoustic scene. More specifically, it is expected that each source will be located close to a grid point in at least one of the set of shifted grids. The sparse solutions corresponding to the set of shifted grids are combined to obtain the source location estimates. The estimated source positions are used as side information to obtain the original source signals.

Joint wideband source localization and acquisition based on a grid-shift approach

The disclosure relates to an apparatus for determining a quality score (MOS) for an audio signal sample, the apparatus comprising: an extractor configured to extract a feature vector from the audio signal sample, wherein the feature vector comprises a plurality of feature values and wherein each feature value is associated to a different feature of the feature vector; a pre-processor configured to pre-process a feature value of the feature vector based on a cumulative distribution function associated to the feature represented by the feature value to obtain a pre-processed feature value; and a processor configured to implement a neural network and to determine the quality score (MOS) for the audio signal sample based on the pre-processed feature value and a set of neural network parameters for the neural network associated to the cumulative distribution function.

Devices and methods for evaluating speech quality

A sound processing node is provided for an arrangement of sound processing nodes configured to receive a plurality of sound signals. The sound processing node comprises a processor configured to generate an output signal based on the plurality of sound signals weighted by a plurality of beamforming weights. The processor is configured to adaptively determine the plurality of beamforming weights on the basis of an adaptive linearly constrained minimum variance beamformer using a transformed version of a least mean squares formulation of a constrained gradient descent approach. The transformed version of the least mean squares formulation of the constrained gradient descent approach is based on a transformation of the least mean squares formulation of the constrained gradient descent approach to the dual domain.

Sound processing node of an arrangement of sound processing nodes

Techniques of performing linear acoustic echo cancellation performing a phase correction operation on the estimate of the echo signal based on a clock drift between a capture of an input microphone signal and a playout of a loudspeaker signal. Along these lines, the existence of the clock drift, i.e., a small difference in the sampling rates of the input microphone signal and the loudspeaker signal, can cause processing circuitry in a device configured to perform LAEC operations to generate a filter based on the magnitudes of the short-term Fourier transforms (STFTs) of the input microphone signal and the loudspeaker signal. Such a filter is real-valued and results in a positive estimate of the acoustic echo signal included in the input microphone signal. The phase of this estimate may then be aligned with the phase of the input microphone signal.

Echo cancellation for keyword spotting

A method includes: receiving a representation of a soundfield, the representation characterizing the soundfield around a point in space; decomposing the received representation into independent signals; and encoding the independent signals, wherein a quantization noise for any of the independent signals has a common spatial profile with the independent signal.

Coding of a soundfield representation

Growing community of inventors

Lower Hutt, New Zealand

Willem Bastiaan Kleijn