.net – the difference between System.Speech.Recognition and Microsoft.Speech.Recognition


There are two similar namespaces and assemblies for speech recognition in .NET. I’m trying to understand the differences and when it is appropriate to use one or the other.

There is System.Speech.Recognition from the assembly System.Speech (in System.Speech.dll). System.Speech.dll is a core DLL in the .NET Framework class library 3.0 and later

There is also Microsoft.Speech.Recognition from the assembly Microsoft.Speech (in microsoft.speech.dll). Microsoft.Speech.dll is part of the UCMA 2.0 SDK

I find the docs confusing and I have the following questions:

System.Speech.Recognition says it is for "The Windows Desktop Speech Technology", does this mean it cannot be used on a server OS or cannot be used for high scale applications?

The UCMA 2.0 Speech SDK ( http://msdn.microsoft.com/en-us/library/dd266409%28v=office.13%29.aspx ) says that it requires Microsoft Office Communications Server 2007 R2 as a prerequisite. However, I’ve been told at conferences and meetings that if I do not require OCS features like presence and workflow I can use the UCMA 2.0 Speech API without OCS. Is this true?

If I’m building a simple recognition app for a server application (say I wanted to automatically transcribe voice mails) and I don’t need features of OCS, what are the differences between the two APIs?

Best Solution

The short answer is that Microsoft.Speech.Recognition uses the Server version of SAPI, while System.Speech.Recognition uses the Desktop version of SAPI.

The APIs are mostly the same, but the underlying engines are different. Typically, the Server engine is designed to accept telephone-quality audio for command & control applications; the Desktop engine is designed to accept higher-quality audio for both command & control and dictation applications.

You can use System.Speech.Recognition on a server OS, but it's not designed to scale nearly as well as Microsoft.Speech.Recognition.

The differences are that the Server engine won't need training, and will work with lower-quality audio, but will have a lower recognition quality than the Desktop engine.