idnavid / spkr_diarization

Alveo Speaker Diarization project - Navid's version

Speaker Diarization component for Alveo project. The role of the diarization tool is to segment long speech files into smaller chunks. The output labels will be used as a benchmark for human transcribers.

Navid Shokouhi July 2017

Packages:

• Spro (for feature extraction)
• AudioSeg (Diarization binaries)
• Python:
  • numpy, scipy

Installation guide:

• Installing Spro 4.0

  cd spro-directory
  ./configure
  make 
  make install 
  

NOTE: when installing on Mac, use Spro 5.0

• Installing AudioSeg:

  cd audioseg-directory
  ./configure --with-spro=[path-to-spro-directory]
  make
  make install
  

diar.py

main module is diar.py, which contains an example script. To load in python, use diar.diarization(root_dir,wavname,ubmname,out_dir), where:

  root_dir: root directory of experiment. 
  wavname: full path to wave file on disk. 
  ubmname: full path to pretrained UBM on disk. 
  out_dir: output directory, to store intermediate files. 

Examples:

A working example for for Austalk data is available in experiments.