Noise robust speech recognition with a switching linear dynamic model
AssigneeMicrosoft Corporation
InventorsJames Droppo, Alejandro Acero
Life of this patent
Estimating the term…
Abstract
A unified, nonlinear, non-stationary, stochastic model is disclosed for estimating and removing effects of background noise on speech cepstra. Generally stated, the model is a union of dynamic system equations for speech and noise, and a model describing how speech and noise are mixed. Embodiments also pertain to related methods for enhancement.
Claims25
A computer-implemented method for enhancing the noise robustness of a speech recognition system, the method comprising:
- applying a switching linear dynamic model to generate a representation of acoustic features;
- and
- utilizing the representation as a component for conducting model based feature enhancement.
The method of claim 3, wherein said utilizing comprises utilizing the probability.
A computer-implemented method for enhancing the noise robustness of a speech recognition system, the method comprising:
- applying a linear dynamic model to generate a representation of speech;
- and
- utilizing the representation as a basis for model based feature enhancement.
A speech recognition system, comprising:
- a switching linear dynamic model configured to be applied to determine a probability of a clean speech feature vector;
- a linear dynamic model configured to be applied to determine a representation of a noise feature;
- an observation model configured to be applied to generate a representation of how speech and noise together produce a noisy observation;
- and
- processing means for utilizing the switching linear dynamic model, the linear dynamic model and the observation model as bases for determining a posterior estimation, the processing means being further configured to apply the posterior estimation as a basis for conducting model based feature enhancement.
Dependencies are inferred from the claim text.