Project Schnitt

• voice activity detector (VAD) algorithm for splitting sound files
• Java Swing GUI for visualization, error analysis and editing
• group project for my MA programming course (Fall, 2015, Kobe University)
• try Demo.exe [5.8MB], Win7 64|32bit tested (no JRE included! -> only runs if Java is installed!)
  • CTRL-A: add selection as hypo interval
  • CTRL-D: delete selection
  • CTRL-N: zoom in
  • CTRL-Q: zoom out

Wave panel

• editable <tt>hypo</tt> tier (green)
• non-editable <tt>target</tt> tier (dark blue)
• automatically updating <tt>eval</tt> tier (red) with VAD errors

Type | Description ------|------------------------------------------------------ TP |true positive; speech detected as speech TN |true negative; silence detected as silence WC |word clipping MSC |mid-speech clipping NDS(1)|noise detected as speech, during silence NDS(2)|noise detected as speech, arching 2 speech activities FEC |front end clipping REC |rear end clipping HEAD |overhead: hypo starts before voice activity TAIL |tail: hypo ends after voice activity ends

Evaluation

The following FDA (more precisely Finite State Transducer) is implemented.