sd

AISHELL4 Speaker Diarization

Usage

In speaker diarization process, we will only use the first channel as the SAD model and the Speaker-Embedding-Extrator do not support the multi-channel wav.

You can find all the process in sd/run.sh and the comments in it.

Here is the main stage:

1. You should first prepare the key file wav.scp for our tools, the scripts local/make_aishell4_test.sh is just a sample, you can use any language (like perl, python or shell) as you like. If you find some naming problems when use, you can replace this step by your own script to prepare the files.
2. You should change your own data path and kaldi-root first in run.sh and path.sh. The script local/do_segmentations.sh is to get the SAD result for the future work. You will find the segments file under the $sad_result_dir.
3. When use the VBx tools for the diarization, you should convert the segments to the .lab. Use scripts/segment_to_lab.sh to change the file format
4. The speaker diarization code needs two stage the speaker-embedding extract and the speaker-embedding cluster. Our baseline use the VBx tools to extract speaker-embeddings. The feature-extractor is inside, you don't have to prepare the feature before. Note our scripts run in a SGE systems sor we sub the extract-embedding jobs to the queue.pl, if you do not have it, try to extract the speaker embeddings one by one. Besides, we recommand you add a exit 1 after the stage 3 to waiting for the extracting process finished.
5. For the speaker-embedding cluster, you can use the run_cluster.sh and the code will make the rttm for each audio in the wav.scp.

Model Download

You need download the model from the path, you should mv the exp to the sd/ and the ResNet101_16kHz to the VBx/models.

Reference

1. kaldi-sad-model
2. VBx