CRSS Speaker Diarization Toolkit (CRSS-SpkrDiar)

About

CRSS-SpkDiar is a C++ based speaker diarization toolkit, built on top of famous open source speech recognition platform of Kaldi. The main objectives of this toolkit are:

Simple integration with Kaldi ASR,
Simple intergration of i-vector modules within Kaldi for Diarization,
Simple intergration of DNN modules within Kaldi for Diarization,
Perform speaker diarization unsupervised/supervised/semi-supervised fashion,
Benchmark on open database (AMI meeting corpus, and Apollo-MCC corpus).

Authors: Chengzhu Yu and Navid Shokouhi.

Current Stage of Development

Completed:

VAD (GMM based)
BIC segmentation (optional)
Bottom-Up Clustering
- BIC distance
- KL divergence
- i-vector cosine distance
- i-vector Mahalanobis
- i-vector PLDA (optional)
Bottom-Up Clustering Using i-vector cosine distance score (CDS)
Interger linear programming (ILP) Clustering

To Be Completed:

VAD based segmentation (viterbi)
Resementation
Evaluations

Furthur Extensions:

DNN speaker embedding features
Interface with Kaldi ASR

Dependencies

Kaldi
GLPK (if only you want to try ILP)

Benchmark Performance

We evaluate our performance on AMI meeting corpus and compare the numbers with those reported in Pycasp from ICSI. Note: To evaluate only the clustering module, the numbers on CRSS-SpkDiar is on top of oracle segmentation. We're currently working to include segmentation.

Session	Pycasp	CRSS-SpkDiar (run2.sh)
IS1000a.Mix-Headset	25.38	12.07
IS1001a.Mix-Headset	32.34	43.64
IS1001b.Mix-Headset	10.57	12.16
IS1001c.Mix-Headset	28.40	6.17
IS1003b.Mix-Headset	34.30	10.56
IS1003d.Mix-Headset	50.75	24.67
IS1006b.Mix-Headset	16.57	7.34
IS1006d.Mix-Headset	53.05	21.56
IS1008a.Mix-Headset	1.65	4.07
IS1008b.Mix-Headset	8.58	3.60
IS1008c.Mix-Headset	9.30	6.36
IS1008d.Mix-Headset	26.27	5.99
Average	24.76%	13.19%

Recommend Projects