C# Speech Recognition

Question

I am making a Smart House Control System right now, and I have a little problem. I was thinking on using Cosmos for a base system, and adding the needed namespace libraries to it, but as the usual System.Speech.Recognition namespace depends too much on Windows Speech API, I have to forget about using it.

So my question is, is there any (free if possible) voice recognition and/or speech speech synthesizer library for C#, what has the following:

• support for multi-language speaking
• extracting text content from speech sample
• synthesizing speech with selectable (or user-written) speech pattern (voice)

A general usage, non-windows dependent library would be the best, and of course, if it was free too.

Answer 1

Check out this project: http://cmusphinx.sourceforge.net/

It's an open source speech recognition project. It is trainable with any language you want plus since its open source you can modify it to suit your needs or expand it.

Answer 2

Voxeo offers developer accounts which you could use to develop a speech powered home automation system. I've interfaced it to my own home automation system for a small subset of the commands my home understands and it works great. You'll need to learn some VoiceXML to use it.

SAPI works OK for voice synthesis; I use SAPI in my system for spoken prompts in the house like a weather forecast that comes over the speakers in the morning when you walk into the bathroom. If Cosmos doesn't allow you to include all the DLLs you need maybe you could create a separate service using SAPI and then use WCF (or other) to communicate between them??

For the related problem of understanding natural language in a typed form I've developed a C# NLP Engine which I hope to be able to make available for non-commercial at some point in the future.

Extracting text from speech without specifying any grammar up-front is a very hard problem and is going to be error prone. Even if you could solve that, you'd still have the problem of trying to understand what they said using NLP. Constructing a grammar that guides the recognizer to the kinds of sentences you want to recognize (like VoiceXML does) is likely to achieve much higher accuracy.