Systems and methods for multi-sourced noise suppression are provided. An example system may receive streams of audio data including a voice signal and noise, the voice signal including a spoken word. The streams of audio data are provided by distributed audio devices. The system can assign weights to the audio streams based at least partially on quality of the audio streams. The weights of audio streams can be determined based on signal-to-noise ratios (SNRs). The system may further process, based on the weights, the audio stream to generate cleaned speech. Each audio device comprises microphone(s) and can be associated with the Internet of Things (IoT), such that the audio devices are Internet of Things devices. The processing can include noise suppression and reduction and echo cancellation. The cleaned speech can be provided to a remote device for further processing which may include Automatic Speech Recognition (ASR).

Multi-Sourced Noise Suppression

A system utilizing two pairs of microphones for noise suppression. Primary and secondary microphones may be positioned closely spaced to each other to provide acoustic signals used to achieve noise cancellation/suppression. An additional, tertiary microphone may be spaced with respect to either the primary microphone or the secondary microphone in a spread-microphone configuration for deriving level cues from audio signals provided by the tertiary and the primary or secondary microphone. The level cues are expressed via a level difference used to determine one or more cluster tracking control signal(s). The level difference-based cluster tracking signals are used to control adaptation of noise suppression. A noise cancelled primary acoustic signal and level difference-based cluster tracking control signals are used during post filtering to adaptively generate a mask to be applied to a speech estimate signal.

Adaptive Noise Reduction Using Level Cues

A signal manipulator for manipulating an audio signal having a transient event may comprise a transient remover (100), a signal processor (110) and a signal inserter (120) for inserting a time portion in a processed audio signal at a signal location where the transient event was removed before processing by said transient remover, so that a manipulated audio signal comprises a transient event not influenced by the processing, whereby the vertical coherence of the transient event is maintained instead of any processing performed in the signal processor (110), which would destroy the vertical coherence of a transient.

Device and method for manipulating an audio signal having a transient event

An encoder capable of reducing the degradation of the quality of the decoded signal in the case of band expansion in which the high band of the spectrum of an input signal is estimated from the low band. In this encoder, a first layer encoding section (202) encodes an input signal and generates first encoded information, a first layer decoding section (203) decodes the first encoded information and generates a first decoded signal, a characteristic judging section (206) analyzes the intensity of the harmonic structure of the input signal and generates harmonic characteristic information representing the analysis result, and a second layer encoding section (207) changes, on the basis of the harmonic characteristic information, the numbers of bits allocated to parameters included in second encoded information created by encoding the difference between the input signal and the first decoded signal before creating the second information .

Encoder, decoder, and encoding method

Described are noise suppression techniques applicable to various systems including automatic speech processing systems in digital audio pre-processing. The noise suppression techniques utilize a machine-learning framework trained on cues pertaining to reference clean and noisy speech signals, and a corresponding synthetic noisy speech signal combining the clean and noisy speech signals. The machine-learning technique is further used to process audio signals in real time by extracting and analyzing cues pertaining to noisy speech to dynamically generate an appropriate gain mask, which may eliminate the noise components from the input audio signal. The audio signal pre-processed in such a manner may be applied to an automatic speech processing engine for corresponding interpretation or processing. The machine-learning technique may enable extraction of cues associated with clean automatic speech processing features, which may be used by the automatic speech processing engine for various automatic speech processing.

Noise suppression for speech processing based on machine-learning mask estimation

This invention presents a voicing determination algorithm for classification of a speech signal segment as voiced or unvoiced. The algorithm is based on a normalized autocorrelation where the length of the window is proportional to the pitch period. The speech segment to be classified is further divided into a number of sub-segments, and the normalized autocorrelation is calculated for each sub-segment if a certain number of the normalized autocorrelation values is above a predetermined threshold, the speech segment is classified as voiced. To improve the performance of the voicing determination algorithm in unvoiced to voiced transients, the normalized autocorrelations of the last sub-segments are emphasized. The performance of the voicing decision algorithm can be enhanced by utilizing also the possible lookahead information.

Method and apparatus for speech coding with voiced/unvoiced determination

of EP1111586This invention presents a voicing determination algorithm for classification of a speech signal segment as voiced or unvoiced. The algorithm is based on a normalised autocorrelation where the length of the window is proportional to the pitch period. The speech segment to be classified is further divided into a number of sub-segments, and the normalised autocorrelation is calculated for each sub-segment. If a certain number of the normalised autocorrelation values is above a predetermined threshold, the speech segment is classified as voiced. To improve the performance of the voicing determination algorithm in unvoiced to voiced transients, the normalised autocorrelations of the last sub-segments are emphasised. The performance of the voicing decision algorithm can be enhanced by utilising also the possible lookahead information.

Method and apparatus for voiced/unvoiced determination

A voicing determination algorithm for classification of a speech signal segment as voiced or unvoiced is based on, for example, a normalised autocorrelation where the length of the window is proportional to the pitch period. The speech segment to be classified is further divided into a number of sub-segments, and the normalised autocorrelation is calculated (302) for each sub-segment. If a certain number (e.g. substantially half) of the normalised autocorrelation values is above a predetermined threshold (304), the speech segment is classified as voiced. To improve the performance of the voicing determination algorithm in unvoiced to voiced transients, the normalised autocorrelations of the last sub-segments are emphasised. The performance of the voicing decision algorithm can be enhanced by utilising also the possible lookahead information. If a certain number of the normalised autocorrelation values of the last sub-segments exceeds another threshold (305), then the speech segment is also classified as voiced.

Samuli Pietila

Papers

Method and apparatus for speech coding with voiced/unvoiced determination

Method and apparatus for voiced/unvoiced determination

Voiced/unvoiced determination for speech coding