d-Ear SR System Description

1. Front end processing

   - 20-ms frame length with 10-ms frame shift
   - Adaptive SS-based denoising
   - Energy based End Point Detection
   - 16-dimension MFCC plus Delta
   - CMS + CVN

2. UBM training (GMM-UBM based framework)

   - Two gender-dependent UBMs, each contains 1024 components
   - Training data comes from NIST SRE'04 set

3. Speaker model training

   - Adapted from gender-dependent UBM with only mean-modified by MAP

4. Score normalization

   - T-Norm with selected imposters from NIST SRE'04 set

5. Multi-speaker processing

   - GLR-based speech segmentation
   - Differential model scores based speaker clustering


** The confidence scores can be interpreted as likelihood ratios.


Hardware Description

   - Single CPU (Intel Pentium-4 at 3.0 GHz)
   - Total memory installed: 1.0 GB


Execution Time (in "Multiple of Real-Time for the data processed")

1. For creating speaker models

   - 1conv4w: 1.0 / 32.89

2. For processing test segments

   - 1conv4w_1conv4w: 1.0 / 45.96
   - 1conv4w_10sec4w: 1.0 / 20.55
   - 1conv4w_1conv2w: 1.0 /  9.95