private static double NoSoundStateEmissionProbability(PitchHmmEmission emission)
{
// 無音 0.3、それ以外 0.7
// 変化したタイミングで入力が来るので、無音でないパターンもそれなりに可能性があると考える
return(Math.Log(emission.IsSilent ? 0.3 : 0.7 / 12.0));
}
public void InputObservation(PitchHmmEmission observation)
{
var states = this._model.States;
var maxLogProbabilities = new double[states.Count];
for (var i = 0; i < maxLogProbabilities.Length; i++)
{
maxLogProbabilities[i] = double.NegativeInfinity;
}
// 初期状態確率 * 遷移確率 * 生成確率 を計算する
foreach (var(fromStateIndex, stateLogProbability) in this._stateLogProbabilities)
{
foreach (var(toStateIndex, transitionLogProbability) in states[fromStateIndex].LogProbabilitiesByOutgoingStateIndexes)
{
var lp = stateLogProbability + transitionLogProbability + states[toStateIndex].EmissionLogProbability(observation);
if (!(lp <= 0.0))
{
throw new Exception("確率が不正な値になりました。"); // NaN チェックも含めるため not <=
}
if (lp > maxLogProbabilities[toStateIndex])
{
maxLogProbabilities[toStateIndex] = lp;
}
}
}
var stateLogProbabilities = this._stateLogProbabilities;
stateLogProbabilities.Clear();
var logSum = double.NegativeInfinity;
var maxP = double.NegativeInfinity; // 最大の確率
var argmaxP = 0; // 最大の確率のときの状態
for (var stateIndex = 0; stateIndex < maxLogProbabilities.Length; stateIndex++)
{
var lp = maxLogProbabilities[stateIndex];
if (double.IsNegativeInfinity(lp))
{
continue;
}
logSum = Special.LogSum(logSum, lp);
stateLogProbabilities.Add(new KeyValuePair <int, double>(stateIndex, lp));
if (lp > maxP)
{
maxP = lp;
argmaxP = stateIndex;
}
}
if (stateLogProbabilities.Count == 0)
{
throw new Exception("すべての状態の確率が 0 になりました。");
}
// 確率が最も高い状態を現在のノートとして認識
this.CurrentNote = states[argmaxP].Value.ReportingNote;
// 次の計算のために確率を均しておく
// 新しい確率 = 確率 / 確率の合計
for (var i = 0; i < stateLogProbabilities.Count; i++)
{
var(stateIndex, lp) = stateLogProbabilities[i];
stateLogProbabilities[i] = new KeyValuePair <int, double>(stateIndex, lp - logSum);
}
}
