/// <summary>
/// Gets the context-free entry point to this state
/// </summary>
/// <returns>the entry point to the state</returns>
/// TODO: ideally we'll look for entry points with no left
// context, but those don't exist yet so we just take
// the first entry point with an SILENCE left context
// note that this assumes that the first node in a grammar has a
// word and that word is a SIL. Not always a valid assumption.
public SentenceHMMState GetEntryPoint()
{
ContextPair cp = ContextPair.Get(UnitContext.Silence, UnitContext.Silence);
List <ISearchState> list = GetEntryPoints(cp);
return(list == null || list.Count == 0 ? null : (SentenceHMMState)list[0]);
}
/**
* /// Attaches the given unit to the given tail, expanding the unit if necessary. If an identical unit is already
* /// attached, then this path is folded into the existing path.
*
* /// @param parent the parent state
* /// @param tail the place to attach the unit to
* /// @param units the set of units
* /// @param which the index into the set of units
* /// @param leftContext the left context for the unit
* /// @param rightContext the right context for the unit
* /// @return the tail of the added unit (or null if the path was folded onto an already expanded path.
*/
private SentenceHMMState AttachUnit(PronunciationState parent,
SentenceHMMState tail, Unit[] units, int which,
UnitContext leftContext, UnitContext rightContext)
{
Unit[] lc = GetLC(leftContext, units, which);
Unit[] rc = GetRC(units, which, rightContext);
UnitContext actualRightContext = UnitContext.Get(rc);
LeftRightContext context = LeftRightContext.Get(lc, rc);
Unit cdUnit = _parent.UnitManager.GetUnit(units[which].Name, units[which].IsFiller, context);
UnitState unitState = new ExtendedUnitState(parent, which, cdUnit);
float logInsertionProbability;
if (unitState.Unit.IsSilence)
{
logInsertionProbability = _parent.LogSilenceInsertionProbability;
}
else if (unitState.Unit.IsFiller)
{
logInsertionProbability = _parent._logFillerInsertionProbability;
}
else if (unitState.GetWhich() == 0)
{
logInsertionProbability = _parent._logWordInsertionProbability;
}
else
{
logInsertionProbability = _parent._logUnitInsertionProbability;
}
// check to see if this state already exists, if so
// branch to it and we are done, otherwise, branch to
// the new state and expand it.
SentenceHMMState existingState = GetExistingState(unitState);
if (existingState != null)
{
AttachState(tail, existingState, LogOne, logInsertionProbability);
// T(" Folding " + existingState);
return(null);
}
else
{
AttachState(tail, unitState, LogOne, logInsertionProbability);
AddStateToCache(unitState);
// T(" Attaching " + unitState);
tail = ExpandUnit(unitState);
// if we are attaching the last state of a word, then
// we add it to the exitPoints table. the exit points
// table is indexed by a ContextPair, consisting of
// the exiting left context and the right context.
if (unitState.IsLast())
{
UnitContext nextLeftContext = GenerateNextLeftContext(leftContext, units[which]);
ContextPair cp = ContextPair.Get(nextLeftContext, actualRightContext);
// T(" Adding to exitPoints " + cp);
AddExitPoint(cp, tail);
}
return(tail);
}
}
/// <summary>
/// Expands each GState into the sentence HMM States
/// </summary>
public void Expand()
{
//this.LogDebug("Item Context: {0} : {1} : {2}",startingContexts.Count, rightContexts.Count, leftContexts.Count);
// for each left context/starting context pair create a list
// of starting states.
foreach (UnitContext leftContext in _leftContexts)
{
foreach (UnitContext startingContext in GetStartingContexts())
{
ContextPair contextPair = ContextPair.Get(leftContext, startingContext);
if (!_entryPoints.ContainsKey(contextPair))
{
_entryPoints.Add(contextPair, new List <ISearchState>());
}
}
}
this.LogDebug("Item entryPoints Count: {0}", _entryPoints.Count);
// if this is a final node don't expand it, just create a
// state and add it to all entry points
if (_node.IsFinalNode)
{
GrammarState gs = new GrammarState(_node);
foreach (List <ISearchState> epList in _entryPoints.Values)
{
epList.Add(gs);
}
}
else if (!_node.IsEmpty)
{
// its a full fledged node with a word
// so expand it. Nodes without words don't need
// to be expanded.
foreach (UnitContext leftContext in _leftContexts)
{
ExpandWord(leftContext);
}
}
else
{
//if the node is empty, populate the set of entry and exit
//points with a branch state. The branch state
// branches to the successor entry points for this
// state
// the exit point should consist of the set of
// incoming left contexts and outgoing right contexts
// the 'entryPoint' table already consists of such
// pairs so we can use that
foreach (var entry in _entryPoints)
{
ContextPair cp = entry.Key;
List <ISearchState> epList = entry.Value;
SentenceHMMState bs = new BranchState(cp.LeftContext.ToString(), cp.RightContext.ToString(), _node.ID);
epList.Add(bs);
AddExitPoint(cp, bs);
}
}
AddEmptyEntryPoints();
}
/**
* /// Expand the pronunciation given the left context
*
* /// @param leftContext the left context
* /// @param pronunciation the pronunciation to expand
* /// @param which unique ID for this pronunciation
*/
// Each GState maintains a list of entry points. This list of
// entry points is used when connecting up the end states of
// one GState to the beginning states in another GState. The
// entry points are tagged by a ContextPair which represents
// the left context upon entering the state (the left context
// of the initial units of the state), and the right context
// of the previous states (corresponding to the starting
// contexts for this state).
//
// When expanding a pronunciation, the following steps are
// taken:
// 1) Get the starting context for the pronunciation.
// This is the set of units that correspond to the start
// of the pronunciation.
//
// 2) Create a new PronunciationState for the
// pronunciation.
//
// 3) Add the PronunciationState to the entry point table
// (a hash table keyed by the ContextPair(LeftContext,
// StartingContext).
//
// 4) Generate the set of context dependent units, using
// the left and right context of the GState as necessary.
// Note that there will be fan out at the end of the
// pronunciation to allow for units with all of the
// various right contexts. The point where the fan-out
// occurs is the (length of the pronunciation - the max
// right context size).
//
// 5) Attach each cd unit to the tree
//
// 6) Expand each cd unit into the set of HMM states
//
// 7) Attach the optional and looping back silence cd
// unit
//
// 8) Collect the leaf states of the tree and add them to
// the exitStates list.
private void ExpandPronunciation(UnitContext leftContext, Pronunciation pronunciation, int which)
{
UnitContext startingContext = GetStartingContext(pronunciation);
// Add the pronunciation state to the entry point list
// (based upon its left and right context)
string pname = "P(" + pronunciation.Word + '[' + leftContext
+ ',' + startingContext + "])-G" + GetNode().ID;
PronunciationState ps = new PronunciationState(pname, pronunciation, which);
_parent.T(" Expanding " + ps.Pronunciation + " for lc " + leftContext);
ContextPair cp = ContextPair.Get(leftContext, startingContext);
var epList = _entryPoints.Get(cp);
if (epList == null)
{
throw new Error("No EP list for context pair " + cp);
}
else
{
epList.Add(ps);
}
Unit[] units = pronunciation.Units;
int fanOutPoint = units.Length - GetRightContextSize();
if (fanOutPoint < 0)
{
fanOutPoint = 0;
}
SentenceHMMState tail = ps;
for (int i = 0; tail != null && i < fanOutPoint; i++)
{
tail = AttachUnit(ps, tail, units, i, leftContext, UnitContext.Empty);
}
SentenceHMMState branchTail = tail;
foreach (UnitContext finalRightContext in _rightContexts)
{
tail = branchTail;
for (int i = fanOutPoint; tail != null && i < units.Length; i++)
{
tail = AttachUnit(ps, tail, units, i, leftContext, finalRightContext);
}
}
}
/**
* /// Adds the set of empty entry points. The list of entry points are tagged with a context pair. The context pair
* /// represent the left context for the state and the starting context for the state, this allows states to be
* /// hooked up properly. However, we may be transitioning from states that have no right hand context (CI units
* /// such as SIL fall into this category). In this case we'd normally have no place to transition to since we add
* /// entry points for each starting context. To make sure that there are entry points for empty contexts if
* /// necessary, we go through the list of entry points and find all left contexts that have a right hand context
* /// size of zero. These entry points will need an entry point with an empty starting context. These entries are
* /// synthesized and added to the the list of entry points.
*/
private void AddEmptyEntryPoints()
{
var emptyEntryPoints = new HashMap <ContextPair, List <ISearchState> >();
foreach (var entry in _entryPoints)
{
ContextPair cp = entry.Key;
if (NeedsEmptyVersion(cp))
{
ContextPair emptyContextPair = ContextPair.Get(cp.LeftContext, UnitContext.Empty);
List <ISearchState> epList = emptyEntryPoints.Get(emptyContextPair);
if (epList == null)
{
epList = new List <ISearchState>();
emptyEntryPoints.Put(emptyContextPair, epList);
}
epList.AddRange(entry.Value);
}
}
_entryPoints.PutAll(emptyEntryPoints);
}