speaker

Speaker Diarization

Usage

For the speaker diarzation track, only the first channel of waves will be used.

The main stage:

1. We use the implementation of Kaldi toolkits. Please install the Kaldi toolkits and conduct ln -s /export/kaldi/utils/ utils and ln -s /export/kaldi/step/ step.
2. Stage 1 is for the data preparation and stage 2 is for the voice detect activity (VAD).
3. When using the VBx toolkits for diarization, please convert the segments file to .lab. Use scripts/segment_to_lab.sh to change the file format.
4. The speaker diarization consists of speaker embedding extraction and speaker embedding clustering. In our baseline system, the VBx toolkit is used to extract the speaker embeddings.
5. For the speaker-embedding cluster, the code will get the hypothesis rttm for each audio in the wav.scp.
6. We obtain the reference rttm through the ground truth transcripts.
7. We use toolkits dscore to get the DER results.

Pre-traiend Model Download

Download the model from the path. Then, move the exp directory to our speaker directory and move the ResNet101_16kHz to speaker/VBx/models.

Reference

1. kaldi-sad-model
2. VBx
3. dscore