Amazon Technologies, Inc.

Hyeong Cheol Kim

Wai Chung Chu

Carlo Murgia

Philip Ryan Hilmes

Robert Ayrapetian

Yuwen Su

A system configured to perform power normalization for voice enhancement. The system may identify active intervals corresponding to voice activity and may selectively amplify the active intervals in order to generate output audio data at a near uniform loudness. The system may determine a variable gain for each of the active intervals based on a desired output loudness and a flatness value, which indicates how much a signal envelope is to be modified. For example, a low flatness value corresponds to no modification, with peak active interval values corresponding to the desired output loudness and lower active intervals being lower than the desired output loudness. In contrast, a high flatness value corresponds to extensive modification, with peak active interval values and lower active interval values both corresponding to the desired output loudness. Thus, individual words may share the same peak power level.

Methods for voice enhancement

A system configured to improve speech quality by performing residual echo suppression (RES). The system may detect when double-talk conditions are present in individual frequency bands during a voice conversation and may determine gain values for the individual frequency bands. The system may determine whether double-talk conditions are present based on a normalized cross power spectral density function in a frequency domain. If double-talk conditions are present in a frequency band or far end energy is low, the system may determine a gain value that passes audio data in the frequency band, whereas if double-talk conditions are not present, the system may determine a gain value that attenuates audio data in the frequency band. The system may determine binary gain values using a decision threshold value or continuous gain values using a mapping function. The system may control an amount of suppression by selecting different mapping functions and/or parameters.

Methods for suppressing residual echo

An echo cancellation system that generates multiple output paths, enabling Automatic Speech Recognition (ASR) processing in parallel with voice communication. For single direction AEC (e.g., ASR processing), the system prioritizes speech from a single user and ignores other speech by selecting a single directional output from a plurality of directional outputs as a first output path. For multi-directional AEC (e.g., voice communication), the system includes all speech by combining the plurality of directional outputs as a second output path. The system may use a weighted sum technique, such that each directional output is represented in the combined output based on a corresponding signal metric, or an equal weighting technique, such that a first group of directional outputs having a higher signal metric may be equally weighted using a first weight while a second group of directional outputs having a lower signal metric may be equally weighted using a second weight.

Growing community of inventors

Irvine, CA, United States of America

Hyeong Cheol Kim