Linear prediction based noise suppression

A device capable of time-domain noise suppression for suppressing a noise signal in a speech signal, said device comprising:

• a signal module including a linear predictor capable of generating an estimate of said speech signal based on a plurality of linear prediction coefficients estimated for said speech signal, wherein said signal module is capable of generating a prediction error estimate based on said plurality of prediction coefficients;
• a noise module including a voice activity detector capable of determining a voice activity in said speech signal and an update noise model element capable of updating a plurality of noise parameters based on said prediction error estimate if said voice activity detector determines no voice activity in said speech signal, and generating an estimate of said noise signal based on said plurality of noise parameters;
• and
• a noise suppression filter derived from said estimate of said noise signal and said estimate of said speech signal, said noise suppression filter capable of receiving said speech signal and generating a clean speech signal estimate;
• wherein said plurality of linear prediction coefficients are associated with a short-term linear predictor indicative of a spectral envelope of said speech signal and a long-term linear predictor indicative of a pitch periodicity of said speech signal, wherein said plurality of noise parameters include a spectral estimate of said noise signal and a residual energy of said noise signal.

A time-domain noise suppression method for suppressing a noise signal in a speech signal, said time-domain noise suppression method comprising:

• estimating a plurality of linear prediction coefficients for said speech signal;
• generating a prediction error estimate based on said plurality of prediction coefficients;
• generating an estimate of said speech signal based on said plurality of linear prediction coefficients;
• using a voice activity detector to determine a voice activity in said speech signal;
• updating a plurality of noise parameters based on said prediction error estimate if said voice activity detector determines no voice activity in said speech signal;
• generating an estimate of said noise signal based on said plurality of noise parameters;
• and
• passing said speech signal through a filter derived from said estimate of said noise signal and said estimate of said speech signal to generate a clean speech signal estimate;
• wherein said plurality of linear prediction coefficients are associated with a short-term linear predictor indicative of a spectral envelope of said speech signal and a long-term linear predictor indicative of a pitch periodicity of said speech signal, wherein said filter is represented by:
• [
• A
• noise
• ⁡
• (
• z
• )
• -
• G
• noise
• ⁢
• A
• LP
• ⁡
• (
• z
• )
• ]
• A
• noise
• ⁡
• (
• z
• )
• ,
• wherein Anoise(z) is a spectral estimate of said noise signal, and GnoiseALP(z) is an estimate of a noise gain.