The official documentation has been moved to http://wiki.sarah.encausse.net/

.
.
.
.
.
.
.
.

In order to have good results you need a good microphone. Here is an article testing Microhpone, Kinect and VoiceTracker II with SARAH.

Improve Recognition

Keep in mind SARAH match voices (from Users, TV, Radio, ...) not music.

Kinect

• Put Kinect in direct line with users.
• Put sound or music source behind Kinect.
• User must speak 20db above sound.

Microphone

• Choose a good Microphone
• Play Windows trainning once (no more)
• Increase device level to 80 (not too much)

SARAH

• Lower confidence in custom.ini
  • The first word "SARAH" should be strong to remove false positive
  • The overall sentence can be lower

custom.ini

; Speech 1st word confidence (aka SARAH)
trigger=0.8

; Speech overall confidence
confidence=0.70

Tweaks

Some SpeechEngine tweaks are exposed in properties to tweak engine behavior. See Microsoft Documentation for more info.

custom.ini

; Speech engines tweaks (in seconds) 
alternate=10
initialSilence=0
babble=0
endSilence=0
endSilenceAmbiguous=0

Debug Recognition

Create a folder /dump. All matching recognition will be dumped in /dump with audio and confidence.

custom.ini

; Print trace logs. Output speech wav in /dump (must exists)    
debug=false

; Log file path (default "${basedir}/${shortdate}.log")
; logfile=${basedir}/${shortdate}.log

; Log2Console port
udpport=9999