EmacsListen FAQ

Setup

• Is the Emacs version important? Yes (but mainly because of cua-mode; if you don't use cua-mode, EmacsListen will likely work with Emacs 20 as well. In fact, it was developed under Emacs 20.)
• Are there any error messages that can safely be ignored? Yes, "Cannot create voice menu: 0x80040626" is made by the vr-mode process. Also, "VR Mode command channel is not open: deactivate-buffer" can be ignored (the command channel code in vr-mode is something that I have touched as little as possible; it probably should be totally removed).

Use

• Dragon NaturallySpeaking correction doesn't highlight the problematic text, why not?

  The text is highlighted in the *EmacsListen* buffer. It is not highlighted in the current buffer, since in general interspersed commands have no representation and that makes it impossible to consistently highlight the spoken utterances in the current buffer.

  Moreover, select-and-say works only for the text that is actually visible (as NaturallySpeaking sees it); this text is the visible contents of the *EmacsListen* buffer. This means that you cannot use the Dragon mechanism in general to select words in the current buffer, but you can still use it for correction of recent commands and text. The ShortTalk command to select a piece of text is choose sorch... (in the forward direction) or choose rorch... (in the backward direction). Of course, all the usual variations apply: choose twoon for choosing the last two words before the cursor, choose inner tair for choosing the inside of a parenthesized expression under the mouse pointer, etc.

  So, remember NaturallySpeaking select-and-say works on the *EmacsListen* buffer.

• Are there any training materials? Please make sure to go through the audio and video demos; then read the quick reference card religiously, it is the key to all of ShortTalk. You may also want to consult the listen/grammar.elg file, which is the formal definition of ShortTalk commands and their semantics. Unfortunately, we have not had resources to prepare a proper tutorial.
• What I say seems to bypass EmacsListen. What to do? If what you say is inserted in your current buffer, but does not appear in the *EmacsListen* buffer, then speech input is routed directly to Emacs (via a default mechanism that ScanSoft provides). Nothing is ever inserted via vr-mode into any buffer, except the *EmacsListen* buffer. When you say "ghin line", NaturallySpeaking simply emits that text into the *EmacsListen* buffer (via vr-mode). (Hook functions in EmacsListen are then invoked to parse the changes in the *EmacsListen* buffer and to produce the desired listen event.) If vr-mode for some reason has gotten confused, then you'll see the words "ghin line" in your current buffer. This means that vr-mode has lost its focus. I suspect what has happened is that you have had two Emacs processes running at the same time. This may sometimes confuse vr-mode, unfortunately. In that case, the Emacs in which you are running *EmacsListen* may not be recognized by vr-mode anymore. If vr-mode loses it, then clicking on the blue bar on *EmacsListen* frame and then clicking on your Emacs frame seem to reset the vr-mode focus.

ShortTalk phonetics

• Are liaisons important? Yes, for example, you should use the following pronunciations:

  word oon	->	wor-DOON
  lat ane	->	la-TANE
  laip ane	->	lai-PANE

Grammar format (.elg)

• What is the syntax for defining EmacsListen associations between spoken utterances and Lisp functions or keyboard macros?

  The format is a BNF notation. Nonterminals are of the form <X>, where is X is a name. Terminals are just names. All rules defined in terms of '=' are expanded in all possible ways according to the definitions of nonterminals. Therefore, the grammar may not contain recursion (cycles). Terminals should be names whose pronunciation can be inferred by the underlying speech engine.

  The ground (all-terminal) expansions of a sentential form alpha are associated with semantic actions. Three ways of defining these associations yield commands that take no argument:

  alpha : function-name;	where interactive function function-name is called.
  alpha : (...);	where s-exp (...) is a list that can be evalulated.
  alpha : [...];	where [...] is a keyboard macro.

  If in a rule ... is inserted between the sentential form and ':', as in alpha ... : function-name, then the rule denotes predicates of the form gamma words, where gamma is a ground expansion of alpha and words is natural language. Sometimes, all of following words do not become part of the command argument, as in the snex command (these commands are those defined by the variable listen-predicate-one-argument-command). Moreover, logic requires that certain commands be recognized when embedded within words. For example, snex tair obviously should produce an XML tag pair whose name is the identifier under the mouse cursor) see variable listen-embedded-commands.

• How do I compile the ShortTalk grammar?

  The ShortTalk grammar is defined in the file grammar.elg. The command M-x listen-load-grammar compiles the grammar into two formats: one is grammar.lcm, a file of lists describing the totality of all commands derived from the grammar, and the other is grammar.txt, which contains a subset of commands, each represented as a pair of the form spoken form/written form. It is assumed that the subset contains enough bigrams to cover all the commands (which are usually bigrams or trigram, although 4-grams also are common) (Currently, even the grammar.txt contains around 20,000 commands. This could probably safely be pruned.) The file grammar.lcm is loaded into a hash table when listen-mode is activated. This is again a solution that is much more space intensive than really needed. However, the resulting parser is fast, since it checks each prefix of possible command words through a single hash table lookup.