search

YongyuG / s4d-diarization-gao

Diarization using s4d toolkit and kaldi
Other
5 stars 6 forks source link

Not found library speakin_voice_feats and speakin_ivector #1

Open giappham opened 5 years ago

giappham commented 5 years ago

Hi @YongyuG . How to download 2 libraries speakin_voice_feats and speakin_ivector in gl_bic_ahc_viterbi.py. I tried installing the environment.yml file but the server died. Many thanks.

chrisspen commented 5 years ago

I ran into this problem as well. I'm assuming these are private tools he developed but hasn't published.

@YongyuG Would it be possible for you to publish these? It looks like only speakin_voice_feats is actually used in the code.

chrisspen commented 5 years ago

Nevermind. I reimplemented the missing libraries myself. You can find them in my fork.

I couldn't get the conda install to work. Most of the dependencies aren't used and so old conda spends hours trying to resolve conflicts. However, I just created a Python virtualenv and that was enough to run everything.

This library is probably the fastest I've ever evaluated. It correctly parses the example1.wav file in a few seconds, whereas most take several minutes. However, it suffers from the same problem as most unsupervised methods in that it doesn't have very good accuracy with non-toy data and can't be corrected through training. When I run it on a short podcast snippet containing three people with similar sounding voices speaking closely together, it incorrectly places three speakers in the same segment.