-record
% julius .... -record directory_nameAll the input to Julius and Julian, regardless of its source type (mic/file/network), will be recorded to the specified directory, with filename of the recorded system time. For example, if an input was occured on 14:53.24 on Feb. 22, 2004, the filename will be:
2004.0222.145324.rawThe format is 16bit, monoral, and the sampling rate depends on the source. The file format is big endian, raw (no header) on Julius/Julian earlier than 3.4, and versions later than 3.4 uses WAV format.
Specifying option "-iwsp
" adds a short pause HMM to all the words in dictionary to handle occasional short pause between words. The added pause HMM will be specially handled as a "Phone Context-Free" (skipped as a triphone context of surrounding phones, and the phone itself will also be unchanged by the context phones) and "Skippable" model, to capture the acoustic property of inter-word short pause, like the figure below:
-spmodel
. The default name is "sp".
Further, you can add an extra transition score to the transition of each short pause model, to control the insertion probability. The score can be specified by "-iwsppenalty
". The value over zero will accelerate the insertion, and specifying value below zero will increase rendency to inhibit the insertion. The default is 0.0.
This "-iwsp
" option can be applied to only multi-path version of Julius / Julian.
There are still a few limitation: you should not define a word in which all consisting phones has direct transitions from initial state to final state. It means that the word itself can be skippable, but Julius / Julian cannot deal with such kind of word-level skipping for now. Such kind of word-level skipping should be considered in language model level, in our current implementation.
This capability of dealing direct skip transition is provided by only multi-path version of Julius / Julian.
-iwspword
" in Julius will add a pause word entry to the dictionary. This option may be effective if you are using N-gram language model which does not consider the location of long silence or breath in a sentence. Specifying this option may improve the recognition accuracy.
The default word to be added is:
<UNK> [sp] sp spYou can also specify the entry above by "
-iwspentry
", like this:
% julius .... -iwspword -iwspentry "<UNK> [sp] sil"
-separatescore
" with monophone AM has been fixed.
configure
opotions:
-iwsp
(mutli-path version)
-iwsppenalty value
(mutli-path version)
-iwspword
(for Julius only)
-iwspentry wordentry
(for Julius only)
-spmodel modelname
-record directory