APPARATUS AND METHOD FOR EXTRACTING NOISE-ROBUST THE SPEECH RECOGNITION VECTOR SHARING THE PREPROCESSING STEP USED IN SPEECH CODING

摘要

An apparatus and a method for extracting noise-robust speech feature vectors by sharing a preprocessing step of a speech coder in a distributed speech recognition terminal are provided to share the preprocessing step for speech communication and speech recognition, thereby improving speech recognition performance as consuming little amount of memory of a lower spec terminal and reducing an operation amount. A channel SNR(Signal-to-Noise Rate) estimation module(24) estimates a channel SNR of a speech signal based on a channel energy estimation value calculated by a channel energy estimation module(23) and a background noise energy estimation value calculated by a background noise estimation module(30). A voice metric calculation module(25) calculates a sum of speech metrics on a channel about the speech signal based on the channel SNR estimated by the channel SNR estimation module. A spectral deviation estimation module(26) estimates a spectrum deviation of the speech signal based on the channel energy estimation value calculated in the channel energy estimation module. A noise update decision module(27) gives a noise estimation value update command based on a difference value among the channel energy estimation value, an estimation value for a current power spectrum, and an estimation value for an average long interval power spectrum. A channel SNR modifier(28) modifies the channel SNR estimated by the channel SNR estimation module based on the sum of the voice metrics. A channel gain computation module(29) computes a linear channel gain based on the modified channel SNR and the background noise energy estimation value. A frequency domain filter(31) applies the linear channel gain to a spectrum signal converted by a frequency domain converter.