Cross-domain filtering for audio noise reduction

摘要

An audio-based system may perform automatic noise reduction to enhance speech intelligibility in an audio signal. Described techniques include initially analyzing audio frames in the time domain to identify frames having relatively low power levels. Those frames are then further analyzed in the frequency domain to estimate noise. For example, the initially identified frames may be analyzed at each of multiple frequencies to detect the lowest exhibited power at each of those frequencies. The lowest power values are used as an estimation of noise across the frequency spectrum, and as the basis for calculating a spectral gain for filtering the audio signal in the frequency domain.

主权项

1. A computing device, comprising:
a processor; an audio input; an audio output; memory, accessible by the processor and storing instructions that are executable by the processor to perform acts comprising:
receiving multiple frames of time-domain audio samples at the audio input;identifying frames of the multiple frames having audio levels that are lower than other frames of the multiple frames;calculating frequency-domain spectrums of individual frames of the identified frames;calculating a power spectral density for individual frames of the identified frames based at least in part on the frequency-domain spectrums of the individual frames, wherein individual ones of the power spectral densities indicate power values of a corresponding one of the identified frames at multiple frequency values;smoothing individual ones of the power spectral densities across the multiple frequency values;at individual ones of the multiple frequency values, identifying a minimum of the power values of the smoothed power spectral densities;calculating a spectral gain based at least in part on the identified minimum power values, wherein the spectral gain indicates a gain value for individual ones of the multiple frequency values;smoothing the spectral gain across the multiple frequency values;filtering the frequency-domain spectrums of the multiple frames based at least in part on the spectral gain; andproducing output audio samples at the audio output based at least in part on the filtered frequency-domain spectrums of the multiple frames.