Acoustic Event Detection

1
Acoustic Event Detection

• Alessio Brutti, Christian Zieger
• FBK-irst, Trento, Italy

2
FBK AED system

• FBK AED system is based on HMM
• It operates in two stages
• Front-end conversion of input audio signals in
  acoustic feature vector sequences (MFCC)
• DetectionClassification stage Viterbi on the
  feature vector sequence to determine the most
  likely event sequence.
• Training both CLEAR 2007 development set ITC
  and UPC isolated event databases.

3
Acoustic Event Detection Front-end

• Front-end
• 1 single fixed microphone audio as input
• sampling frequency at 44.1 kHz
• 20 ms analysis window
• 10 ms step
• Hamming window
• Feature vector is composed by 38 elements
• 12 MFCC first derivatives second derivatives
• First and second derivative of the log energy

4
Acoustic Event Detection HMM

• HMM topology
• 3 states, left-to-right
• Mixtures of 128 Gaussians
• Diagonal covariance matrix
• A model is trained for each event to be detected
• Speech and silence models are added to reject
  speech and silence sequences
• Audio signal of each seminar is converted into a
  feature vector sequence
• The Viterbi algorithm is applied to the feature
  vector sequence to detect the optimum event
  sequence

5
SI and SD systems

• Two systems are implemented
• Site Independent (SI)
• all data available in the development set are
  used for training
• Site Dependent (SD)
• a set of models are trained for each room
• ITC ITC isolated event database ITC meetings
• UPC UPC isolated event database UPC meetings
• UKA ITC isolated event database UKA meetings
• AIT ITC isolated event database AIT meetings
• In the detection stage models are selected on the
  prior information about site

6
Results
7

• Thank you, questions?