/**
* Reads the FST file in the given path, and creates the nodes in the FST file.
*
* @param path the path of the FST file to read
* @return the highest ID of all nodes
* @throws java.io.IOException
*/
private int CreateNodes(String path)
{
ExtendedStreamTokenizer tok = new ExtendedStreamTokenizer(path, true);
int maxNodeId = 0;
while (!tok.IsEOF())
{
tok.Skipwhite();
String token = tok.GetString();
if (token == null)
{
break;
}
else if (token.Equals("T"))
{
tok.GetInt("src id"); // toss source node
int id = tok.GetInt("dest id"); // dest node numb
if (id > maxNodeId)
{
maxNodeId = id;
}
String word1 = tok.GetString(); // get word
if (word1 == null)
{
continue;
}
String word2 = tok.GetString(); // get word
tok.GetString(); // toss probability
String nodeName = "G" + id;
GrammarNode node = _nodes.Get(nodeName);
if (node == null)
{
if (word2.Equals(","))
{
node = CreateGrammarNode(id, false);
}
else
{
node = CreateGrammarNode(id, word2);
}
_nodes.Put(nodeName, node);
}
else
{
if (!word2.Equals(","))
{
/*
* if (!word2.equals(getWord(node))) {
* System.out.println(node + ": " + word2 + ' ' + getWord(node)); }
*/
Debug.Assert(word2.Equals(GetWord(node)));
}
}
}
}
tok.Close();
return(maxNodeId);
}
/**
* /// Loads the sphinx3 density file, a set of density arrays are created and
* /// placed in the given pool.
* ///
* /// @param useCDUnits
* /// if true, loads also the context dependent units
* /// @param inputStream
* /// the open input stream to use
* /// @param path
* /// the path to a density file
* /// @throws FileNotFoundException
* /// if a file cannot be found
* /// @throws IOException
* /// if an error occurs while loading the data
*/
protected void LoadHMMPool(Boolean useCDUnits, Stream inputStream,
string path)
{
var est = new ExtendedStreamTokenizer(inputStream,
'#', false);
this.LogInfo("Loading HMM file from: " + path);
est.ExpectString(ModelVersion);
var numBase = est.GetInt("numBase");
est.ExpectString("n_base");
var numTri = est.GetInt("numTri");
est.ExpectString("n_tri");
var numStateMap = est.GetInt("numStateMap");
est.ExpectString("n_state_map");
var numTiedState = est.GetInt("numTiedState");
est.ExpectString("n_tied_state");
var numContextIndependentTiedState = est
.GetInt("numContextIndependentTiedState");
est.ExpectString("n_tied_ci_state");
var numTiedTransitionMatrices = est.GetInt("numTiedTransitionMatrices");
est.ExpectString("n_tied_tmat");
var numStatePerHMM = numStateMap / (numTri + numBase);
Debug.Assert(numTiedState == MixtureWeightsPool.StatesNum);
Debug.Assert(numTiedTransitionMatrices == MatrixPool.Size);
// Load the base phones
for (var i = 0; i < numBase; i++)
{
var name = est.GetString();
var left = est.GetString();
var right = est.GetString();
var position = est.GetString();
var attribute = est.GetString();
var tmat = est.GetInt("tmat");
var stid = new int[numStatePerHMM - 1];
for (var j = 0; j < numStatePerHMM - 1; j++)
{
stid[j] = est.GetInt("j");
Debug.Assert(stid[j] >= 0 && stid[j] < numContextIndependentTiedState);
}
est.ExpectString("N");
Debug.Assert(left.Equals("-"));
Debug.Assert(right.Equals("-"));
Debug.Assert(position.Equals("-"));
Debug.Assert(tmat < numTiedTransitionMatrices);
var unit = _unitManager.GetUnit(name, attribute.Equals(Filler));
ContextIndependentUnits.Put(unit.Name, unit);
//this.LogInfo("Loaded " + unit.ToString());
// The first filler
if (unit.IsFiller && unit.Name.Equals(SilenceCiphone))
{
unit = UnitManager.Silence;
}
var transitionMatrix = MatrixPool.Get(tmat);
var ss = GetSenoneSequence(stid);
IHMM hmm = new SenoneHMM(unit, ss, transitionMatrix, GetHMMPosition(position));
HmmManager.Put(hmm);
}
if (HmmManager.Get(HMMPosition.Undefined, UnitManager.Silence) == null)
{
throw new IOException("Could not find SIL unit in acoustic model");
}
// Load the context dependent phones. If the useCDUnits
// property is false, the CD phones will not be created, but
// the values still need to be read in from the file.
var lastUnitName = "";
Unit lastUnit = null;
int[] lastStid = null;
SenoneSequence lastSenoneSequence = null;
for (var i = 0; i < numTri; i++)
{
var name = est.GetString();
var left = est.GetString();
var right = est.GetString();
var position = est.GetString();
var attribute = est.GetString();
var tmat = est.GetInt("tmat");
var stid = new int[numStatePerHMM - 1];
for (var j = 0; j < numStatePerHMM - 1; j++)
{
stid[j] = est.GetInt("j");
Debug.Assert(stid[j] >= numContextIndependentTiedState &&
stid[j] < numTiedState);
}
est.ExpectString("N");
Debug.Assert(!left.Equals("-"));
Debug.Assert(!right.Equals("-"));
Debug.Assert(!position.Equals("-"));
Debug.Assert(attribute.Equals("n/a"));
Debug.Assert(tmat < numTiedTransitionMatrices);
if (useCDUnits)
{
Unit unit;
var unitName = (name + ' ' + left + ' ' + right);
if (unitName.Equals(lastUnitName))
{
unit = lastUnit;
}
else
{
var leftContext = new Unit[1];
leftContext[0] = ContextIndependentUnits.Get(left);
var rightContext = new Unit[1];
rightContext[0] = ContextIndependentUnits.Get(right);
Context context = LeftRightContext.Get(leftContext,
rightContext);
unit = _unitManager.GetUnit(name, false, context);
}
lastUnitName = unitName;
lastUnit = unit;
//this.LogInfo("Loaded " + unit.ToString());
var transitionMatrix = MatrixPool.Get(tmat);
var ss = lastSenoneSequence;
if (ss == null || !SameSenoneSequence(stid, lastStid))
{
ss = GetSenoneSequence(stid);
}
lastSenoneSequence = ss;
lastStid = stid;
IHMM hmm = new SenoneHMM(unit, ss, transitionMatrix, GetHMMPosition(position));
HmmManager.Put(hmm);
}
}
est.Close();
}
private void GetSenoneToCIPhone()
{
var inputStream = GetDataStream(Path.Combine(Location.Path, Model));
if (inputStream == null)
{
throw new IOException("can't find modelDef " + Model);
}
var est = new ExtendedStreamTokenizer(inputStream, '#', false);
this.LogInfo("Loading HMM file from: " + Model);
est.ExpectString(ModelVersion);
_numBase = est.GetInt("numBase");
est.ExpectString("n_base");
var numTri = est.GetInt("numTri");
est.ExpectString("n_tri");
var numStateMap = est.GetInt("numStateMap");
est.ExpectString("n_state_map");
var numTiedState = est.GetInt("numTiedState");
est.ExpectString("n_tied_state");
Senone2Ci = new int[numTiedState];
est.GetInt("numContextIndependentTiedState");
est.ExpectString("n_tied_ci_state");
var numTiedTransitionMatrices = est.GetInt("numTiedTransitionMatrices");
est.ExpectString("n_tied_tmat");
var numStatePerHMM = numStateMap / (numTri + _numBase);
Debug.Assert(numTiedState == MixtureWeightsPool.StatesNum);
Debug.Assert(numTiedTransitionMatrices == MatrixPool.Size);
// Load the base phones
for (var i = 0; i < _numBase + numTri; i++)
{
//TODO name this magic const somehow
for (var j = 0; j < 5; j++)
{
est.GetString();
}
var tmat = est.GetInt("tmat");
for (var j = 0; j < numStatePerHMM - 1; j++)
{
Senone2Ci[est.GetInt("j")] = tmat;
}
est.ExpectString("N");
Debug.Assert(tmat < numTiedTransitionMatrices);
}
est.Close();
}
/// <summary>
/// Returns the next Data object, which is the mel cepstrum of the input frame. However, it can also be other Data objects like DataStartSignal.
/// </summary>
/// <returns>
/// The next available Data object, returns null if no Data object is available.
/// </returns>
/// <exception cref="System.Exception">
/// IOException closing cepstrum stream
/// or
/// IOException reading from cepstrum stream
/// </exception>
public override IData GetData()
{
IData data;
if (_curPoint == -1)
{
data = new DataStartSignal(_sampleRate);
_curPoint++;
}
else if (_curPoint == _numPoints)
{
if (_numPoints > 0)
{
_firstSampleNumber =
(_firstSampleNumber - _frameShift + _frameSize - 1);
}
// send a DataEndSignal
var numberFrames = _curPoint / _cepstrumLength;
var totalSamples = (numberFrames - 1) * _frameShift + _frameSize;
var duration = (long)
((totalSamples / (double)_sampleRate) * 1000.0);
data = new DataEndSignal(duration);
try {
if (_binary)
{
_binaryStream.Close();
}
else
{
_est.Close();
}
_curPoint++;
} catch (IOException ioe) {
throw new Exception("IOException closing cepstrum stream", ioe);
}
}
else if (_curPoint > _numPoints)
{
data = null;
}
else
{
var vectorData = new double[_cepstrumLength];
for (var i = 0; i < _cepstrumLength; i++)
{
try {
if (_binary)
{
if (_bigEndian)
{
vectorData[i] = _binaryStream.ReadFloat();
}
else
{
vectorData[i] = Utilities.ReadLittleEndianFloat(_binaryStream);
}
}
else
{
vectorData[i] = _est.GetFloat("cepstrum data");
}
_curPoint++;
} catch (IOException ioe) {
throw new Exception("IOException reading from cepstrum stream", ioe);
}
}
// System.out.println("Read: " + curPoint);
data = new DoubleData
(vectorData, _sampleRate, _firstSampleNumber);
_firstSampleNumber += _frameShift;
// System.out.println(data);
}
return(data);
}
/// <summary>
/// Creates the grammar.
/// </summary>
/// <returns>The initial node for the grammar.</returns>
protected override GrammarNode CreateGrammar()
{
GrammarNode initialNode = null;
GrammarNode finalNode = null;
// first pass create the FST nodes
int maxNodeId = CreateNodes(_path);
// create the final node:
finalNode = CreateGrammarNode(++maxNodeId, IDictionary.SilenceSpelling);
finalNode.SetFinalNode(true);
// replace each word node with a pair of nodes, which
// consists of the word node and a new dummy end node, which is
// for adding null or backoff transitions
maxNodeId = ExpandWordNodes(maxNodeId);
ExtendedStreamTokenizer tok = new ExtendedStreamTokenizer(_path, true);
// Second pass, add all of the arcs
while (!tok.IsEOF())
{
String token;
tok.Skipwhite();
token = tok.GetString();
// System.out.println(token);
if (token == null)
{
break;
}
else if (token.Equals("I"))
{
Debug.Assert(initialNode == null);
int initialID = tok.GetInt("initial ID");
String nodeName = "G" + initialID;
// TODO: FlatLinguist requires the initial grammar node
// to contain a single silence. We'll do that for now,
// but once the FlatLinguist is fixed, this should be
// returned to its former method of creating an empty
// initial grammar node
// initialNode = createGrammarNode(initialID, false);
initialNode = CreateGrammarNode(initialID, IDictionary.SilenceSpelling);
_nodes.Put(nodeName, initialNode);
// optionally add a silence node
if (_addInitialSilenceNode)
{
GrammarNode silenceNode = CreateGrammarNode(++maxNodeId, IDictionary.SilenceSpelling);
initialNode.Add(silenceNode, LogMath.LogOne);
silenceNode.Add(initialNode, LogMath.LogOne);
}
}
else if (token.Equals("T"))
{
int thisID = tok.GetInt("this id");
int nextID = tok.GetInt("next id");
GrammarNode thisNode = Get(thisID);
GrammarNode nextNode = Get(nextID);
// if the source node is an FSTGrammarNode, we want
// to join the endNode to the destination node
if (HasEndNode(thisNode))
{
thisNode = GetEndNode(thisNode);
}
float lnProb = 0f; // negative natural log
String output = tok.GetString();
if (output == null || output.Equals(","))
{
// these are epsilon (meaning backoff) transitions
if (output != null && output.Equals(","))
{
tok.GetString(); // skip the word
lnProb = tok.GetFloat("probability");
}
// if the destination node has been expanded
// we actually want to add the backoff transition
// to the endNode
if (HasEndNode(nextNode))
{
nextNode = GetEndNode(nextNode);
}
}
else
{
String word = tok.GetString(); // skip words
lnProb = tok.GetFloat("probability");
if (_ignoreUnknownTransitions && word.Equals("<unknown>"))
{
continue;
}
/*
* System.out.println(nextNode + ": " + output);
*/
Debug.Assert(HasWord(nextNode));
}
thisNode.Add(nextNode, ConvertProbability(lnProb));
}
else if (token.Equals("F"))
{
int thisID = tok.GetInt("this id");
float lnProb = tok.GetFloat("probability");
GrammarNode thisNode = Get(thisID);
GrammarNode nextNode = finalNode;
if (HasEndNode(thisNode))
{
thisNode = GetEndNode(thisNode);
}
thisNode.Add(nextNode, ConvertProbability(lnProb));
}
}
tok.Close();
Debug.Assert(initialNode != null);
return(initialNode);
}
