Meeting Recorder: Portable Speech Recognition

Meeting Recorder Notes

Vector Processing

A vector processor (such as IRAM) allows certain types of regular computations (e.g. multimedia processing) to be performed very efficiently both in terms of execution time and energy consumption. Many components of speech recognition are highly regular, and may be efficiently executed on a vector architecture.

Spontaneous Speech

Normal speech between people, as opposed to read speech, prepared speech, or speech to a computer. Spontaneous speech contains many disfluencies, including false starts, restarts, repeated words, filled pauses (uh, um), etc. Speech recognition on spontaneous speech is much harder than dictation.

Function Words

The words in an utterance that do not carry semantic meaning. Informally, the 200 or so most common English words, including "a", "the", "an", "of", etc. Function words tend to be monosyllabic and unstressed in English. Speech recognition systems typically perform poorly on content words. Information retrieval systems usually remove all content words, both from the corpus and the query.

Content Words

The words in an utterance that carry the semantic content or meaning. Informally, all the non-function words. Content words tend to longer than function words, and also tend to be stressed in English. As a result, speech recognition systems usually perform better on content words than on function words.

[ Back ]