Abstract: Disclosed are systems, methods, and other implementations for noise suppression, including a method that includes obtaining a sound signal sample, determining a noise reduction profile, from a plurality of noise reduction profiles, for processing the obtained sound signal sample, and processing the sound signal sample with a machine learning system to produce a noise suppressed signal. The machine learning system implements (executes) a single machine learning model trained to controllably suppress noise in input sound signals according to the plurality of noise reduction profiles. The processing of the sound signal sample is performed according to the determined noise reduction profile.

Abstract: A method for applying a watermark signal to a speech signal to prevent unauthorized use of speech signals, the method may include receiving an original speech signal; determining a corresponding spectrogram of the original speech signal; selecting a phase sequence of fixed frame length and uniform distribution; and generating an encoded watermark signal based on the corresponding spectrogram and phase sequence.

Abstract: A siren detector identifies those frequencies at which a siren is active and a localizer localizes the siren based on those frequencies.

Abstract: A method detects presence of a multi-tone siren type in an acoustic signal. The multi-tone siren type is associated with one or more siren patterns, where each siren pattern includes a number of time patterns at corresponding frequencies. The method includes processing a number of frequency components of a frequency domain representation of the acoustic signal over time to determine a corresponding plurality of values. That processing includes determining, for each frequency component, a value characterizing a presence of a time pattern associated with at least one siren pattern. The method also includes processing the values according to the siren patterns to determine a detection result indicating whether the multi-tone siren type is present in the acoustic signal.

Abstract: A system and method for detecting multi-tone sirens despite environmental noises that may be present obtains a microphone input signal, applies, in real time, a time-frequency analysis to the microphone input signal to determine a time-frequency representation, provides at least one multi-tone model that has a plurality of tone duration patterns, performs multi-tone siren detection on the time-frequency representation, the detection based on the at least one multi-tone model and factoring of doppler shifts, and generates a detection result that can be used in systems for automated vehicles.

Abstract: A method, computer program product, and computer system for transforming, by a computing device, a speech signal into a speech signal representation. A regression deep neural network may be trained with a cost function to minimize a mean squared error between actual values of the speech signal representation and estimated values of the speech signal representation, wherein the cost function may include one or more discriminative terms. Bandwidth of the speech signal may be extended by extending the speech signal representation of the speech signal using the regression deep neural network trained with the cost function that includes the one or more discriminative terms.

Abstract: A system and method for detecting multi-tone sirens despite environmental noises that may be present obtains a microphone input signal, applies, in real time, a time-frequency analysis to the microphone input signal to determine a time-frequency representation, provides at least one multi-tone model that has a plurality of tone duration patterns, performs multi-tone siren detection on the time-frequency representation, the detection based on the at least one multi-tone model and factoring of doppler shifts, and generates a detection result that can be used in systems for automated vehicles.

Abstract: A method, computer program product, and computer system for receiving, by a computing device, an input signal. A first power spectral density estimate may be generated for a linear reverberant component associated with the input signal. A second power spectral density estimate may be generated for a non-linear reverberant component associated with the input signal. A power spectral density estimate may be generated by combining the first power spectral density estimate for the linear reverberant component and the second power spectral density estimate for the non-linear reverberant component. One or more parameters for at least one of the linear reverberant component and the non-linear reverberant component may be updated. One or more undesired signal components in an output signal resulting from the input signal may be reduced via residual echo suppression based upon, at least in part, updating the one or more parameters.

Abstract: A method for residual echo suppression is provided. Embodiments may include receiving an original reference signal and applying a distortion function to the original reference signal to generate a second signal. Embodiments may include generating a non-linear signal from the distortion function that does not include linear components of the original reference signal. Embodiments may also include calculating a residual echo power of a linear component and a non-linear component, wherein the linear component is based upon the original reference signal and the non-linear component is based upon the non-linear signal. Embodiments may further include applying a room model to each of the original reference signal and the non-linear signal and estimating a power associated with the original reference signal and the non-linear signal. Embodiments may include calculating a combined echo power estimate as a weighted sum of a weighted original reference signal power and a weighted non-linear signal power.

Abstract: A method, computer program product, and computer system for transforming, by a computing device, a speech signal into a speech signal representation. A regression deep neural network may be trained with a cost function to minimize a mean squared error between actual values of the speech signal representation and estimated values of the speech signal representation, wherein the cost function may include one or more discriminative terms. Bandwidth of the speech signal may be extended by extending the speech signal representation of the speech signal using the regression deep neural network trained with the cost function that includes the one or more discriminative terms.

Abstract: A system and method for speech enhancement of a portable electronic device. Embodiments may include receiving an audio signal at a portable electronic device having a first microphone and a second microphone. Embodiments may also include receiving an input from a proximity detector associated with the portable electronic device and controlling a processing component associated with at least one of the first microphone and the second microphone based upon, at least in part, the input from the proximity detector.

Abstract: A method, computer program product, and computer system for receiving, by a computing device, an input signal. A first power spectral density estimate may be generated for a linear reverberant component associated with the input signal. A second power spectral density estimate may be generated for a non-linear reverberant component associated with the input signal. A power spectral density estimate may be generated by combining the first power spectral density estimate for the linear reverberant component and the second power spectral density estimate for the non-linear reverberant component. One or more parameters for at least one of the linear reverberant component and the non-linear reverberant component may be updated. One or more undesired signal components in an output signal resulting from the input signal may be reduced via residual echo suppression based upon, at least in part, updating the one or more parameters.

Abstract: A system and method for speech enhancement of a portable electronic device. Embodiments may include receiving an audio signal at a portable electronic device having a first microphone and a second microphone. Embodiments may also include receiving an input from a proximity detector associated with the portable electronic device and controlling a processing component associated with at least one of the first microphone and the second microphone based upon, at least in part, the input from the proximity detector.

Abstract: Methods and apparatus for dynamically suppressing low frequency non-speech audio events, such as road bumps, without suppressing speech formants. In exemplary embodiments of the invention, maximum powers in first and second windows are computed and used to determine whether dampening should be applied, and if so, to what extent.

Abstract: Methods and apparatus for dynamically suppressing low frequency non-speech audio events, such as road bumps, without suppressing speech formants. In exemplary embodiments of the invention, maximum powers in first and second windows are computed and used to determine whether dampening should be applied, and if so, to what extent.

Abstract: The invention relates to an assistance system (1, 25) for providing support in situation-dependent planning and/or guidance tasks of a controlled system (2), comprising a state detection unit (3) for detecting at least one state (5, 5a) of the controlled system (2), an acoustic receiving unit (9, 27), which is designed to receive acoustic voice signals of voice communication (24) between at least two persons (21, 22), and a voice processing unit (7), which is designed to detect voice information (10) regarding the controlled system (2) from the received acoustic voice signals, wherein the state detection unit (3) is designed to adapt the detected state (5) and/or a predicted state (5a) of the controlled system (2) that can be derived from the current state, according to the detected voice information (10).