static Collection<List<String>> extractCorrectSentenceList(
List<SpeechNBestList> speechNBestLists) {
Collection<List<String>> correctSentences = new ArrayList<List<String>>();
for (SpeechNBestList speechNBestList : speechNBestLists) {
correctSentences.add(speechNBestList.getCorrectSentence());
}
return correctSentences;
}
static double calculateWordErrorRateUpperBound(List<SpeechNBestList> speechNBestLists) {
double totalDistance = 0.0;
double totalWords = 0.0;
EditDistance editDistance = new EditDistance();
for (SpeechNBestList speechNBestList : speechNBestLists) {
List<String> correctSentence = speechNBestList.getCorrectSentence();
double worstDistance = Double.NEGATIVE_INFINITY;
for (List<String> guess : speechNBestList.getNBestSentences()) {
double distance = editDistance.getDistance(correctSentence, guess);
if (distance > worstDistance) worstDistance = distance;
}
totalDistance += worstDistance;
totalWords += correctSentence.size();
}
return totalDistance / totalWords;
}
static double calculateWordErrorRateRandomChoice(List<SpeechNBestList> speechNBestLists) {
double totalDistance = 0.0;
double totalWords = 0.0;
EditDistance editDistance = new EditDistance();
for (SpeechNBestList speechNBestList : speechNBestLists) {
List<String> correctSentence = speechNBestList.getCorrectSentence();
double sumDistance = 0.0;
double numGuesses = 0.0;
for (List<String> guess : speechNBestList.getNBestSentences()) {
double distance = editDistance.getDistance(correctSentence, guess);
sumDistance += distance;
numGuesses += 1.0;
}
totalDistance += sumDistance / numGuesses;
totalWords += correctSentence.size();
}
return totalDistance / totalWords;
}
static double calculateWordErrorRate(
LanguageModel languageModel, List<SpeechNBestList> speechNBestLists, boolean verbose) {
double totalDistance = 0.0;
double totalWords = 0.0;
EditDistance editDistance = new EditDistance();
for (SpeechNBestList speechNBestList : speechNBestLists) {
List<String> correctSentence = speechNBestList.getCorrectSentence();
List<String> bestGuess = null;
double bestScore = Double.NEGATIVE_INFINITY;
double numWithBestScores = 0.0;
double distanceForBestScores = 0.0;
for (List<String> guess : speechNBestList.getNBestSentences()) {
double score =
Math.log(languageModel.getSentenceProbability(guess))
+ (speechNBestList.getAcousticScore(guess) / 16.0);
double distance = editDistance.getDistance(correctSentence, guess);
if (score == bestScore) {
numWithBestScores += 1.0;
distanceForBestScores += distance;
}
if (score > bestScore || bestGuess == null) {
bestScore = score;
bestGuess = guess;
distanceForBestScores = distance;
numWithBestScores = 1.0;
}
}
// double distance = editDistance.getDistance(correctSentence,
// bestGuess);
totalDistance += distanceForBestScores / numWithBestScores;
totalWords += correctSentence.size();
if (verbose) {
if (distanceForBestScores > 0.0) {
System.out.println();
displayHypothesis("GUESS:", bestGuess, speechNBestList, languageModel);
displayHypothesis("GOLD: ", correctSentence, speechNBestList, languageModel);
// System.out.println("GOLD: "+correctSentence);
}
}
}
return totalDistance / totalWords;
}
private static void displayHypothesis(
String prefix,
List<String> guess,
SpeechNBestList speechNBestList,
LanguageModel languageModel) {
double acoustic = speechNBestList.getAcousticScore(guess) / 16.0;
double language = Math.log(languageModel.getSentenceProbability(guess));
System.out.println(
prefix
+ " AM: "
+ nf.format(acoustic)
+ " LM: "
+ nf.format(language)
+ " Total: "
+ nf.format(acoustic + language)
+ " "
+ guess);
}