Saturday, March 9, 2013

Speech recognition settings tests


Always in the quest for better performance, I decided to fiddle around with the various audio settings available for speech recognition and see if any of them yielded better results.

Base system (unless noted otherwise):

Core i5 760 no overclock, 4 gig ram, Gigabyte GA-P55-USB3 motherboard
Win 7
Dragon NaturallySpeaking 11.0
Andrea PureAudio USB soundpod external soundcard
VXI microphone

Methodology:


A new user with only the initial training
Same text: National Geographic magazine November 2005 page 1: Who Knew?
Change a variable
Compare number of misrecognitions on each variable change
If uncertain, repeat


Raw data:


# of misrecognitions
Nc= noise cancellation
Nonc= no noice cancellation

30 Micro boost


NC
25
Nonc
30
Nc vxi
21
Nc+ LB
32+

0 boost


Regular
41
A NC
40
Vxi
23
NC Vxi
21

Vxi dif Lmodel
18
+NC
29


Andrea
27
Vxi
14
Vxi Nc
15

Other text


Vxi
28
Vxi nc
26
Vxi dif Lmodel
22
Vxi dif Lmodel NC
23

  

VXI Within a virtual machine


Out
15
VM
15

Vxi dif Lmodel 10db boost


Vxi
24
Vxi Nc
21

vxi

 

Dragon trained
10
Win7 Speech recognition no training
40

Acoustic adaptation


Pre optimized
13
Post
44
Post round#2
14


Conclusions:

  • A sound pod is better than integrated sound
  • Microphone boost, software noise cancellation and surround settings are best left alone
  • The VXI microphone is way better than my old Andrea NC8
  • From the get go, Dragon NaturallySpeaking is better than Windows speech recognition
  • There is no impact on using it within a virtual machine
  • Training the software and correcting misrecognitions works, though I've had mixed results with the acoustic adaptation feature
  • Just from personal experience, there are limits to training: some misrecognitions will always remain

0 comments:

Post a Comment