RESEARCH

Acoustic & Speech

  1. Home
  2. RESEARCH
  3. Acoustic & Speech

Investigates various Speech signal processing schemes for acoustic modeling so that more robust speech recognition can be achieved. Our aim is to perform the state-of-art research providing effective means for achieving:

 

Abnormal Acoustic Event Localization and Recognition

작성자 관리자 날짜 2021-06-03 16:06:47 조회수 345
 

Abnormal acoustic event localization and detection system

Contents

  1. Introduction
  2. Main algorithm and principle
  3. Application demo

 

1. Introduction
    - This is multiple abnormal acoustic event localization and detection system using 5 microphones. There is a time delay corresponding to the direction of sound between microphones. We estimate the time delay and find the direction of the abnormal sounds. After finding the direction of sound, we use acoustic beamforming technique to separate the input signal into each direction, one by one. Then we classify the each of separated acoustic signals into one of the pre-defined abnormal events. Even when multiple acoustic events occurs simultaneously, we can recognize and localize each of them

2. Main algorithm and principle
    - Relevant algorithms

  1. Steered Response Power – Phase Transform
    1. After computing Generalized Cross Correlation of microphone pairs, use minimum filter to find peaks of steered response power
    2. The peaks are candidates for direction of acoustic event
    3. Detect the directions of acoustic events by applying threshold to the candidates
  2. Beamforming technique for source separation
    1. Compensate the arrival time difference between input signals of each microphone. The arrival time difference is computed using detected direction of acoustic event. This step makes input signals of each microphone corresponding to the detected direction have same phase in time-frequency domain
    2. By applying phase error based masking, enhance the input signal from the detected direction
  3. Abnormal event classification
    1. Train Gaussian mixture models with MFCC (Mel-frequency Cepstrum Coefficient) features of abnormal acoustic events (man yelling, woman screaming, baby crying, glass breaking, siren, skidding) and normal acoustic event (babble)
    2. Extract MFCC feature of input signal and evaluate it using likelihood classification

    - System structure

3. Application demo

 

 

목록

댓글 (0)

등록된 댓글이 없습니다.
작성 권한이 없습니다.