Apple Inc.

Jason Wung

Joshua D Atkins

Mehrez Souden

Ramin Pishehvar

Ante Jukic

Daniele Giacobello

Ashrith Deshpande

Martin E Johnson

Jonathan D Sheaffer

Juha O Merimaa

Symeon Delikaris Manias

Peter A Raffensperger

Sarmad Aziz Malik

Raghavendra S Prabhu

Feipeng Li

Ramin Pichevar

Disclosed is a multi-task machine learning model such as a time-domain deep neural network (DNN) that jointly generate an enhanced target speech signal and target audio parameters from a mixed signal of target speech and interference signal. The DNN may encode the mixed signal, determine masks used to jointly estimate the target signal and the target audio parameters based on the encoded mixed signal, apply the mask to separate the target speech from the interference signal to jointly estimate the target signal and the target audio parameters, and decode the masked features to enhance the target speech signal and to estimate the target audio parameters. The target audio parameters may include a voice activity detection (VAD) flag of the target speech. The DNN may leverage multi-channel audio signal and multi-modal signals such as video signals of the target speaker to improve the robustness of the enhanced target speech signal.

End-to-end time-domain multitask learning for ML-based speech enhancement

A plurality of microphone signals can be captured with a plurality of microphones of the device. One or more echo dominant audio signals can be determined based on a pick-up beam directed towards one or more speakers of a playback device. Sound that is emitted from the one or more speakers and sensed by the plurality of microphones can be removed from plurality of microphone signals, by using the one or more echo dominant audio signals as a reference, resulting in clean audio.

Spatially informed acoustic echo cancelation

A device implementing a system for processing speech in an audio signal includes at least one processor configured to receive an audio signal corresponding to at least one microphone of a device, and to determine, using a first model, a first probability that a speech source is present in the audio signal. The at least one processor is further configured to determine, using a second model, a second probability that an estimated location of a source of the audio signal corresponds to an expected position of a user of the device, and to determine a likelihood that the audio signal corresponds to the user of the device based on the first and second probabilities.

Spatially informed audio signal processing for user speech

An echo canceller is disclosed in which audio signals of the playback content received by one or more of the microphones from a loudspeaker of the device may be used as the playback reference signals to estimate the echo signals of the playback content received by a target microphone for echo cancellation. The echo canceller may estimate the transfer function between a reference microphone and the target microphone based on the playback reference signal of the reference microphone and the signal of the target microphone. To mitigate near-end speech cancellation at the target microphone, the echo canceller may compute a mask to distinguish between target microphone audio signals that are echo-signal dominant and near-end speech dominant. The echo canceller may use the mask to adaptively update the transfer function or to modify the playback reference signal used by the transfer function to estimate the echo signals of the playback content.

Echo cancellation using a subset of multiple microphones as reference channels

Processing input audio channels for generating spatial audio can include receiving a plurality of microphone signals that capture a sound field. Each microphone signal can be transformed into a frequency domain signal. From each frequency domain signal, a direct component and a diffuse component can be extracted. The direct component can be processed with a parametric renderer. The diffuse component can be processed with a linear renderer. The components can be combined, resulting in a spatial audio output. The levels of the components can be adjusted to match a direct to diffuse ratio (DDR) of the output with the DDR of the captured sound field. Other aspects are also described and claimed.

Processing of audio signals for spatial audio

A number of features are extracted from a current frame of a multi-channel speech pickup and from side information that is a linear echo estimate, a diffuse signal component, or a noise estimate of the multi-channel speech pickup. A DNN-based speech presence probability is produced for the current frame, where the SPP value is produced in response to the extracted features being input to the DNN. The DNN-based SPP value is applied to configure a multi-channel filter whose input is the multi-channel speech pickup and whose output is a single audio signal. In one aspect, the system is designed to run online, at low enough latency for real time applications such voice trigger detection. Other aspects are also described and claimed.

Deep learning driven multi-channel filtering for speech enhancement

A digital speech enhancement system that performs a specific chain of digital signal processing operations upon multi-channel sound pick up, to result in a single, enhanced speech signal. The operations are designed to be computationally less complex yet as a whole yield an enhanced speech signal that produces accurate voice trigger detection and low word error rates by an automatic speech recognizer. The constituent operations or components of the system have been chosen so that the overall system is robust to changing acoustic conditions, and can deliver the enhanced speech signal with low enough latency so that the system can be used online (enabling real-time, voice trigger detection and streaming ASR.) Other embodiments are also described and claimed.

Multi-channel speech signal enhancement for robust voice trigger detection and automatic speech recognition

Method for performing speech enhancement using a Deep Neural Network (DNN)-based signal starts with training DNN offline by exciting a microphone using target training signal that includes signal approximation of clean speech. Loudspeaker is driven with a reference signal and outputs loudspeaker signal. Microphone then generates microphone signal based on at least one of: near-end speaker signal, ambient noise signal, or loudspeaker signal. Acoustic-echo-canceller (AEC) generates AEC echo-cancelled signal based on reference signal and microphone signal. Loudspeaker signal estimator generates estimated loudspeaker signal based on microphone signal and AEC echo-cancelled signal. DNN receives microphone signal, reference signal, AEC echo-cancelled signal, and estimated loudspeaker signal and generates a speech reference signal that includes signal statistics for residual echo or for noise. Noise suppressor generates a clean speech signal by suppressing noise or residual echo in the microphone signal based on speech reference signal. Other embodiments are described.

System and method for performing speech enhancement using a deep neural network-based signal

An echo canceller can be arranged to receive an input signal and to receive a reference signal. The echo canceller can subtract a linear component of the reference signal from the input signal. A noise suppressor can suppress non-linear effects of the reference signal in the input signal in correspondence with a large number of selectable parameters. Such suppression can be provided on a frequency-by-frequency basis, with a unique set of tunable parameters selected for each frequency. A degree of suppression provided by the noise suppressor can correspond to an estimate of residual echo remaining after the one or more linear components of the reference signal have been subtracted from the input signal, to an estimated double-talk probability, and to an estimated signal-to-noise ratio of near-end speech in the input signal for each respective frequency. A speech recognizer can receive a processed input signal from the noise suppressor.

Growing community of inventors

Los Angeles, CA, United States of America

Jason Wung