This is a Python implementation of the Denoising Autoencoder approach that we proposed for the first Multi-target speaker detection and identification Challenge Evaluation (MCE 2018, https://www.mce2018.org ).
The basic idea is to train a Denoising Autoencoder to map each individual input ivector to the mean of all ivectors from that speaker. The aim of this DAE is to compensate for inter-session variability and increase the discriminative power of the ivectors.
You can find our system description for the MCE 2018 challenge here.
The task for the MCE 2018 Evaluation was to detect if a given speech segment belongs to any of the speakers in a blacklist. The challenge is divided into two related subtasks: Top-S detection, i.e. detecting if the segment belongs to any of the blacklist speakers; and Top-1 detection, i.e. detecting which specific blacklist speaker (if any) is speaking in the segment. The data was generated from real call center user-agent telephone conversations. Instead of raw audio data, organizers processed the original data and provided 600-dimensional ivectors. This way, no special signal processing knowledge was needed to enter the evaluation. More details can be found on the evaluation plan.
The dataset can be found at:
https://www.kaggle.com/kagglesre/blacklist-speakers-dataset
After download, extract the files to data folder.
Our training script shows how a very simple DAE can bring a very nice improvement over the baseline. If you run
python mce2018_dae_tst.py
you should get results like these:
Dev set score using train set :
Top S detector EER is 2.40%
Top 1 detector EER is 9.50% (Total confusion error is 343)
Test set score using train set:
Top S detector EER is 6.83%
Top 1 detector EER is 12.42% (Total confusion error is 411)
Test set score using train + dev set:
Top S detector EER is 5.69%
Top 1 detector EER is 8.90% (Total confusion error is 257)
Note that these results do not match our official submission to the challenge, were we obtained Top-S EER: 4.33%, Top-1 EER: 6.11%, since our final system was a bit more complex including Probabilistic Linear Discriminant Analysis (PLDA) scoring and Symmetric Normalization (S-Norm).
- Numpy
- Scikit-learn
- Pandas
- Keras
The code should run on both Python 2 and 3.