public TrainedLTS(AllophoneSet aPhonSet, CART predictionTree) {
this.allophoneSet = aPhonSet;
this.tree = predictionTree;
this.featureDefinition = tree.getFeatureDefinition();
this.indexPredictedFeature = featureDefinition.getFeatureIndex(PREDICTED_STRING_FEATURENAME);
Properties props = tree.getProperties();
if (props == null)
throw new IllegalArgumentException("Prediction tree does not contain properties");
convertToLowercase = Boolean.parseBoolean(props.getProperty("lowercase"));
context = Integer.parseInt(props.getProperty("context"));
}
/**
* Convenience method to load tree from an inputstream
*
* @param treeStream
* @throws IOException
*/
public void loadTree(InputStream treeStream) throws IOException, MaryConfigurationException {
MaryCARTReader cartReader = new MaryCARTReader();
this.tree = cartReader.loadFromStream(treeStream);
this.featureDefinition = tree.getFeatureDefinition();
this.indexPredictedFeature = featureDefinition.getFeatureIndex(PREDICTED_STRING_FEATURENAME);
this.convertToLowercase = false;
Properties props = tree.getProperties();
if (props == null)
throw new IllegalArgumentException("Prediction tree does not contain properties");
convertToLowercase = Boolean.parseBoolean(props.getProperty("lowercase"));
context = Integer.parseInt(props.getProperty("context"));
}
/**
* Create a utterance model list from feature vectors predicted from elements.
*
* @param predictFromElements elements from MaryXML from where to get feature vectors.
* @return Utterance model um containing state durations and pdfs already searched on the trees to
* generate F0.
* @throws MaryConfigurationException if error searching in HMM trees.
*/
private HTSUttModel createUttModel(List<Element> predictFromElements)
throws MaryConfigurationException {
int i, k, s, t, mstate, frame, durInFrames, durStateInFrames, numVoicedInModel;
HTSModel m;
List<Element> predictorElements = predictFromElements;
List<Target> predictorTargets = getTargets(predictorElements);
FeatureVector fv;
HTSUttModel um = new HTSUttModel();
FeatureDefinition feaDef = htsData.getFeatureDefinition();
float duration;
double diffdurOld = 0.0;
double diffdurNew = 0.0;
float f0s[] = null;
try {
// (1) Predict the values
for (i = 0; i < predictorTargets.size(); i++) {
fv = predictorTargets.get(i).getFeatureVector();
Element e = predictFromElements.get(i);
um.addUttModel(new HTSModel(cart.getNumStates()));
m = um.getUttModel(i);
/* this function also sets the phone name, the phone between - and + */
m.setPhoneName(fv.getFeatureAsString(feaDef.getFeatureIndex("phone"), feaDef));
/* Check if context-dependent gv (gv without sil) */
if (htsData.getUseContextDependentGV()) {
if (m.getPhoneName().contentEquals("_")) m.setGvSwitch(false);
}
/* increment number of models in utterance model */
um.setNumModel(um.getNumModel() + 1);
/* update number of states */
um.setNumState(um.getNumState() + cart.getNumStates());
// get the duration from the element
duration = Integer.parseInt(e.getAttribute("d")) * 0.001f; // in sec.
// distribute the duration (in frames) among the five states, here it is done the same
// amount for each state
durInFrames = (int) (duration / fperiodsec);
durStateInFrames = (int) (durInFrames / cart.getNumStates());
m.setTotalDur(0); // reset to set new value according to duration
for (s = 0; s < cart.getNumStates(); s++) {
m.setDur(s, durStateInFrames);
m.setTotalDur(m.getTotalDur() + m.getDur(s));
}
um.setTotalFrame(um.getTotalFrame() + m.getTotalDur());
System.out.format(
"createUttModel: duration=%.3f sec. durInFrames=%d durStateInFrames=%d m.getTotalDur()=%d\n",
duration, durInFrames, durStateInFrames, m.getTotalDur());
/* Find pdf for LF0, this function sets the pdf for each state.
* and determines, according to the HMM models, whether the states are voiced or unvoiced, (it can be possible that some states are voiced
* and some unvoiced).*/
cart.searchLf0InCartTree(m, fv, feaDef, htsData.getUV());
for (mstate = 0; mstate < cart.getNumStates(); mstate++) {
for (frame = 0; frame < m.getDur(mstate); frame++)
if (m.getVoiced(mstate)) um.setLf0Frame(um.getLf0Frame() + 1);
}
}
return um;
} catch (Exception e) {
throw new MaryConfigurationException(
"Error searching in tree when creating utterance model. ", e);
}
}
/**
* Predict durations and state durations from predictFromElements and apply durations to
* applyToElements. A utterance model is created that contains the predicted state durations.
*
* @param predictFromElements elements to predict from
* @param applyToElements elements to apply predicted durations
* @return HTSUttModel a utterance model
* @throws MaryConfigurationException if error searching in HMM trees.
*/
private HTSUttModel predictAndSetDuration(
List<Element> predictFromElements, List<Element> applyToElements)
throws MaryConfigurationException {
List<Element> predictorElements = predictFromElements;
List<Target> predictorTargets = getTargets(predictorElements);
FeatureVector fv = null;
HTSUttModel um = new HTSUttModel();
FeatureDefinition feaDef = htsData.getFeatureDefinition();
double diffdurOld = 0.0;
double diffdurNew = 0.0;
String durAttributeName = "d";
try {
// (1) Predict the values
for (int i = 0; i < predictorTargets.size(); i++) {
fv = predictorTargets.get(i).getFeatureVector();
um.addUttModel(new HTSModel(cart.getNumStates()));
HTSModel m = um.getUttModel(i);
/* this function also sets the phone name, the phone between - and + */
m.setPhoneName(fv.getFeatureAsString(feaDef.getFeatureIndex("phone"), feaDef));
/* Check if context-dependent gv (gv without sil) */
if (htsData.getUseContextDependentGV()) {
if (m.getPhoneName().contentEquals("_")) m.setGvSwitch(false);
}
/* increment number of models in utterance model */
um.setNumModel(um.getNumModel() + 1);
/* update number of states */
um.setNumState(um.getNumState() + cart.getNumStates());
// Estimate state duration from state duration model (Gaussian)
diffdurNew = cart.searchDurInCartTree(m, fv, htsData, diffdurOld);
diffdurOld = diffdurNew;
double duration = m.getTotalDur() * fperiodsec; // in seconds
um.setTotalFrame(um.getTotalFrame() + m.getTotalDur());
// System.out.format("HMMModel: phone=%s duration=%.3f sec. m.getTotalDur()=%d\n",
// m.getPhoneName(), duration, m.getTotalDur());
/* Find pdf for LF0, this function sets the pdf for each state.
* and determines, according to the HMM models, whether the states are voiced or unvoiced, (it can be possible that some states are voiced
* and some unvoiced).*/
cart.searchLf0InCartTree(m, fv, feaDef, htsData.getUV());
for (int mstate = 0; mstate < cart.getNumStates(); mstate++) {
for (int frame = 0; frame < m.getDur(mstate); frame++)
if (m.getVoiced(mstate)) um.setLf0Frame(um.getLf0Frame() + 1);
}
// set the value in elements
Element element = applyToElements.get(i);
// "evaluate" pseudo XPath syntax:
// TODO this needs to be extended to take into account targetAttributeNames like "foo/@bar",
// which would add the
// bar attribute to the foo child of this element, creating the child if not already
// present...
if (durAttributeName.startsWith("@")) {
durAttributeName = durAttributeName.replaceFirst("@", "");
}
String formattedTargetValue = String.format(targetAttributeFormat, duration);
// System.out.println("HMMModel: formattedTargetValue = " + formattedTargetValue);
// if the attribute already exists for this element, append targetValue:
if (element.hasAttribute(durAttributeName)) {
formattedTargetValue =
element.getAttribute(durAttributeName) + " " + formattedTargetValue;
}
// set the new attribute value:
element.setAttribute(durAttributeName, formattedTargetValue);
}
return um;
} catch (Exception e) {
throw new MaryConfigurationException("Error searching in tree when predicting duration. ", e);
}
}