Acoustic Model Adaptation Using Splines
AssigneeMicrosoft Corporation
InventorsMichael Lewis Seltzer, Kaustubh Prakash Kalgaonkar, Alejandro Acero
Life of this patent
Estimating the term…
Abstract
Described is a technology by which a speech recognizer is adapted to perform in noisy environments using linear spline interpolation to approximate the nonlinear relationship between clean speech, noise, and noisy speech. Linear spline parameters that minimize the error the between predicted noisy features and actual noisy features are learned from training data, along with variance data that reflect regression errors. Also described is compensating for linear channel distortion and updating noise and channel parameters during speech recognition decoding.
Claims20
In a computing environment, a method performed on at least one processor, comprising, adapting a speech recognizer to a noisy environment, including approximating a nonlinear relationship between speech, noise and noisy speech parameters by using a spline function to obtain adaptation parameters of a spline model, inputting noise and noisy speech, selecting a set of adaptation parameters from the spline model based upon the noise input, and using the adaptation parameters to adapt speech parameters for use in recognizing the noisy speech.
In a computing environment, a system comprising, a speech recognizer, including a speech model having speech parameters learned from speech training data, and a spline model having adaptation parameters learned from the speech training data and noise samples, the spline model including a subset of the adaptation parameters for each frequency filter corresponding to a spline function for that frequency filter that approximates a nonlinear relationship between speech, noise and noisy speech parameters, the speech recognizer configured to receive noisy speech input, to select selected adaptation parameters based upon sampled noise, and to adapt the speech parameters using the selected adaptation parameters.
One or more computer-readable media having computer-executable instructions, which when executed perform steps, comprising:
- (a) inputting a noisy utterance;
- (b) obtaining initial estimates of noise and channel parameters;
- (c) computing spline weights and transformation parameters based upon the estimates;
- (d) adapting speech parameters into adapted parameters based upon the spline weights and transformation parameters computed at step (c);
- (e) decoding the utterance based upon the adapted parameters;
- and
- (f) outputting data corresponding to the decoded utterance.
Dependencies are inferred from the claim text.