Nyílt Forráskód Techblog: augusztus 2009

SIMON: things to improve

More feedback during recognition and training (the new "Sam" tool can help a lot)
Measuring noise levels
Mass train of words
Quickstart and wizards

Simon quickstart package (with Julius quickstart package)

SIMON vs. Windows 7 speech recognition

This comparison is based on 7 days of testing. Hopefully Simon development will follow the "Ubuntu" way. In my opinion Simon GUI can be a little confusing - a quickstart wizard and quick help on the sidebars would be greatly appreciated by end-users.

SIMON

Distribution: OSS, cross platform
Start year: 2007
Ease of use: quite hard - no demo or quickstart package (with Voxforge stuff - at least English). You should read the manual thoroughly. Don't expect quick results. If you want to test the core, just download the Julius quickstart package: http://www.voxforge.org/
Flexibility: Custom commands (programs, places, text macros, etc.)
Accuracy: After a week, it's still not sufficent (false positives, activation to wrong words or noise)

TIP: If recognition is below e.g. 85%, SIMON should drop the input and do nothing

Windows 7

Speech recognition in Windows 7 (trained for 20 minutes with reading specific text aloud + completed the tutorial that also trains the engine)
Distribution: Closed source, Windows integrated application
Start year: 1993
Ease of use: easy
Flexibility: built-in commands, Windows Speech Recognition Macros recommended
http://www.microsoft.com/downloads/details.aspx?FamilyID=fad62198-220c-4717-b044-829ae4f7c125&displaylang=en
Accuracy: about 80% (without rare words and if you speak 4-5 words at a time)

http://en.wikipedia.org/wiki/Windows_Speech_Recognition

Notable incident

The use of Windows Speech Recognition during a demonstration of Windows Vista at a Microsoft Financial Analyst Meeting on July 27, 2006, resulted in a well-publicized and embarrassing incident. The software failed to function correctly initially, resulting in an unintended output of "Dear aunt, let's set so double the killer delete select all".^[5] ^[6] ^[7] A developer with Vista's speech recognition team later explained that Windows Speech Recognition's failure to function properly during the demonstration was the result of a bug in the volume control feature, which caused the application to pick up extra noise that affected its performance.^[8] ^[9] The software bug was fixed by Microsoft prior to the release of Vista to the general public.^[8]

Testing SIMON speech recogniton software

I have tested the Simon listens speech recogniton system for more than a week and I'll write down my experiences in this post. I've also had the chance to test Windows 7 speech recognition, so I can make a rough comparison.

Why we need speech recognition? For the majority it's just for convenience, but for many handicapped people it means the only way to interact with their environment. Or it can make their life much easier. A worthy goal to fight for.

My project goal: to create a speaker (in)dependent engine that can recognise up to 50 Hungarian words with high accuracy.

Possible sources of errors?

1. WORSE RECOGNITION RATE IF: Recording one word several times in one file?
2. No indication of sound input! I must keep Audacity running all times. A little stay-on top window would help that could indicate the current sound strength (like Windows), a switch and the recognised word.

How to avoid common sources of errors?

* Use external sound card or USB microphone to avoid white noise - OK!
* Use a good quality microphone (I used a Logitech S 7500 webcam, it has echo cancellation and provides an almost noise free recording in silence)