bingo-todd / gmm_localize

Training

Train set

• BRIR

  Roomsim_Campbell

• Sound source

  TIMIT

  • Train set

    1. 8 mic positions

    2. 21 target source positions (-50~50)

    3. 10 interfering source position($$\pm 5,\pm 10,\pm 20,\pm 30,\pm 40$$)

    4. 3 SNR (0,10,20 dB)

  sentence = 5040

  sentence_mic_pos = 630

Test set

To obtain reliable models, GMMs are train with features selected based on 4 criterias:

1. VAD (before frequency analysis)

   Select frames of which energy exceed the threshold (energe maximum - 40)

2. Target is dominant

the signals of the left and the right ear were added prior to energy computaion

$$ \begin{equation} \begin{aligned} SNR = \frac{\sum\left(s^{tar}_L(t)+s^{tar}_R(t)\right)^2}{\sum\left(s^{inter}_L(t)+s^{inter}_R(t)\right)^2} \end{aligned} \end{equation} $$

3. L,R channels are correlated

   Frame with binaural cross-correlation maximum exceed the threshold (0.3)

4. ITD estimations are within (-1,1)ms

Cues

e.g.

Training

Example: sound source in the front (azimuth=0), training result of 32 frequency bands

Evaluation

Comparable result in reference paper (hightlighted)

Illustration of azimuth estimation

#source	-50	0	50
1
2
3
4