/**
* Counts all the tokens in the active list (and displays them). This is an
* expensive operation.
*/
protected void showTokenCount()
{
if (logger.IsInfoEnabled)
{
HashSet <Token> tokenSet = new HashSet <Token>();
var tokens = JavaToCs.GetTokenCollection(activeList);
foreach (Token token in tokens)
{
var tok = token;
while (tok != null)
{
tokenSet.Add(tok);
tok = tok.getPredecessor();
}
}
logger.Info("Token Lattice size: " + tokenSet.Count);
tokenSet = new HashSet <Token>();
foreach (Token token in resultList)
{
var tok = token;
while (tok != null)
{
tokenSet.Add(tok);
tok = tok.getPredecessor();
}
}
logger.Info("Result Lattice size: " + tokenSet.Count);
}
}
/**
* Because the growBranches() is called although no data is left after the
* last speech frame, the ordering of the active-list might depend on the
* transition probabilities and (penalty-scores) only. Therefore we need to
* undo the last grow-step up to final states or the last emitting state in
* order to fix the list.
*
* @return newly created list
*/
protected ActiveList undoLastGrowStep()
{
ActiveList fixedList = activeList.newInstance();
var tokens = JavaToCs.GetTokenCollection(activeList);
foreach (Token token in tokens)
{
Token curToken = token.getPredecessor();
// remove the final states that are not the real final ones because
// they're just hide prior final tokens:
while (curToken.getPredecessor() != null &&
((curToken.isFinal() &&
curToken.getPredecessor() != null && !curToken
.getPredecessor().isFinal()) ||
(curToken.isEmitting() && curToken.getData() == null) // the
// so
// long
// not
// scored
// tokens
|| (!curToken.isFinal() && !curToken.isEmitting())))
{
curToken = curToken.getPredecessor();
}
fixedList.add(curToken);
}
return(fixedList);
}
/**
* Goes through the active list of tokens and expands each token, finding
* the set of successor tokens until all the successor tokens are emitting
* tokens.
*/
protected void growBranches()
{
int mapSize = activeList.size() * 10;
if (mapSize == 0)
{
mapSize = 1;
}
growTimer.start();
bestTokenMap = new Dictionary <SearchState, Token>(mapSize);
ActiveList oldActiveList = activeList;
resultList = new List <Token>();
activeList = activeListFactory.newInstance();
threshold = oldActiveList.getBeamThreshold();
wordThreshold = oldActiveList.getBestScore() + logRelativeWordBeamWidth;
var tokens = JavaToCs.GetTokenCollection(oldActiveList);
foreach (Token token in tokens)
{
collectSuccessorTokens(token);
}
growTimer.stop();
if (logger.IsInfoEnabled)
{
int hmms = activeList.size();
totalHmms += hmms;
logger.Info("Frame: " + currentFrameNumber + " Hmms: " + hmms
+ " total " + totalHmms);
}
}
//@Override
public override void newProperties(PropertySheet ps)
{
base.newProperties(ps);
//logger = ps.getLogger();
name = ps.getInstanceName();
logMath = (LogMath)ps.getComponent(PROP_LOG_MATH);
linguist = (edu.cmu.sphinx.linguist.Linguist)ps.getComponent(PROP_LINGUIST);
pruner = (Pruner)ps.getComponent(PROP_PRUNER);
scorer = (AcousticScorer)ps.getComponent(PROP_SCORER);
activeListFactory = (ActiveListFactory)ps
.getComponent(PROP_ACTIVE_LIST_FACTORY);
showTokenCountP = JavaToCs.ConvertBool(ps.getBoolean(PROP_SHOW_TOKEN_COUNT));
double relativeWordBeamWidth = ps.getDouble(PROP_RELATIVE_WORD_BEAM_WIDTH);
growSkipInterval = ps.getInt(PROP_GROW_SKIP_INTERVAL);
wantEntryPruning = JavaToCs.ConvertBool(ps.getBoolean(PROP_WANT_ENTRY_PRUNING));
logRelativeWordBeamWidth = logMath.linearToLog(relativeWordBeamWidth);
this.keepAllTokens = true;
}
