The front end can be broken in several simple pieces as depicted in the following illustration
Examining the illustration left to right, top to bottom, the front end first reads in raw data via the Data Framer. This data can arrive by a stream, a file, or any other means. In addition, while the data is typically audio, there can be a number of simultaneous data sources, such as both video and audio. The Data Framer packetizes the data into annotated Data Frames. These Data Frames contain information about the data packets, including information such as if the data is the beginning or end of a segment.

The front end passes the Data Frames to a series of Data Processors. The Data Processors perform successive modifications to the Data Frames, such as automatic gain control, noise cancellation, down/up sampling, and preemphasis.

Once the preprocessing of a Data Frame is complete, the front end passes the Data Frame to a Feature Extractor. The Feature Extractor extracts the feature(s) necessary for the decoder to do its work. For audio, this typically involves obtaining cepstral and delta cepstral information, but can be anything the decoder accepts. Furthermore, the resulting Feature is not necessarily restricted to one data type. For example, the Feature may contain information for both audio and video.

The front end then passes Feature frames to to a series of Feature Processors. The Feature Processors may perform a number of operations including end pointing, noise cancellation, and cepstral mean calculation.

When the front end has completed processing a Feature, it passes it to the decoder.

Last edited Oct 12, 2014 at 5:06 PM by DxN, version 3