This code reflects the work described in the InterSpeech'2016 paper on Automatic Dialect Detection in Arabic Broadcast Speech.

It also contains a baseline system for the VarDial'2017 shared task on Arabic Dialect Identification.

• Python (tested with v.2.7.5)
• Multi-class SVM (http://www.cs.cornell.edu/people/tj/svm_light/svm_multiclass.html)

• We provide data for five Arabic dialects: Egyptian (EGY), Levantine (LAV), Gulf (GLF), North African (NOR), and Modern Standard Arabic (MSA).

• The data comes from broadcast news.

VarDial'2017 shared task shared data, and features.

• The baseline for VarDial'2017 is using data/train.vardial2017/ and data/dev.vardial2017/ for training and development default
• For each dialect, there are two features files:
• $dialect.words -- lexical features generated using LVCSR- generated using QCRI MGB-2 submission.
• $dialect.ivec -- i-vector based on bottleneck features, with a fixed length of 400 per utterance.
• wav.lst -- link to the original audio files; WAVE audio, Microsoft PCM, 16 bit, mono 16000 Hz.
• Baseline-- bottleneck iVectors 57.28% accuracy and lexical features 48.43%.

InterSpeech'2016 paper shared data.

• To reproduce the results in InterSpeech'2016, the script should point to data/train.IS2016/ and data/test.IS2016/ for training and testing.
• $dialect.words -- lexical features generated using LVCSR;
• $dialect.ivec -- i-vector based on bottleneck features, with a fixed length of 400 per utterance.
• $dialect.phones -- phoneme sequence from an automatic phoneme recognition system.
• $dialect.phone_duration -- phoneme sequence, and the duration in milliseconds for each phone, e.g., w_030 means phone w for 30 milliseconds.

Run 'run.sh' for an example of the code and the data

• features=phones -- you can use words, phones or ivectors;
• context=6 -- for some features, less context might be enough;
• NOTE 1: The regularization parameters can be optimized for better performance.
• NOTE 2: System combination can be explored as well.

This data and the baseline system are described in this paper:

@inproceedings{ali2016automatic,
  author={Ali, Ahmed and Dehak, Najim and Cardinal, Patrick and Khurana, Sameer and Yella, Sree Harsha and Glass, James and Bell, Peter and Renals, Steve},
  title={Automatic Dialect Detection in Arabic Broadcast Speech},
  booktitle={Interspeech},
  address={San Francisco, CA, USA}
  pages={2934--2938},
  year={2016}
}