Rss Categories

Transcription Word Lists

Reference Number: AA-02493 Views: 2183 0 Rating/ Voters

Important Note

The information here applies to the DNN based ASR Engine which is available in the new LumenVox cloud-native software stack.

A transcription recognition uses statistical language models (SLMs) to help with recognition.  The SLM can be augmented by using a list of words or phrases, in case they suffer poor recognition.  Such words or phrases might not exist in the SLM model, or might have very low probability.  For example, proper names, location names, or drug names would not ordinarily be recognized well, because they are absent or not well represented in the language model.  Adding a word list, will help recognition for these use cases. To make it easy we are using a grammar format to write the word lists, but keep in mind, we use this mechanism to augment the SLM, NOT to augment or fix a grammar.

Ordinarily, to trigger an SLM transcription, you start a recognition with a “dummy” grammar, which is an empty grammar containing the meta tag TRANSCRIPTION_ENGINE set to V2.  This dummy empty grammar is what we use for the word list.  The meta tag LM_MODIFICATION_TYPE with a value of WORDLIST, signals the engine to treat the dummy grammar file contents as a word list.  Below is an example adding the words "anthropomorphize" and "ibuprofen", and the phrases "San Diego", "New York" and "Tom Jones"

<?xml version='1.0'?> <grammar xml:lang="es" version="1.0" root="root" mode="voice"          xmlns=""          tag-format="semantics/1.0"> <meta name="TRANSCRIPTION_ENGINE" content="V2"/> <meta name="LM_MODIFICATION_TYPE" content="WORDLIST"/> <rule id="root" scope="public">   <item> anthropomorphize ibuprofen "San Diego" "New York" "Tom Jones"         </item>   </rule> </grammar>

The WORDLIST mode has two additional parameters, word/phrase probability score multiplier, and word/phrase confidence score boost.  These can be set via meta tags, which are set for the entire list.

The probability score boost raises or lowers the probability the words or phrases are recognized.
The value is a number between -5 and 15. A negative value lowers the probability the word is returned in results.

The Default value is 0.

1<meta name="LM_WORDLIST_PROBABILITY_BOOST" content="0"/>