audio_quality_analyzer.py

"""
音声品質分析モジュール

K-fold CERと実機CERの乖離を理解するための分析ツール
- 音声品質スコア
- 予測CER範囲
- モデルの弱点音素の検出
"""
import numpy as np
import librosa


class AudioQualityAnalyzer:
    """音声品質分析クラス"""

    # モデルの既知の弱点（Phase 7-4の分析から）
    DIFFICULT_PHONEMES = {
        'ʔ': 'glottal_stop',     # 声門閉鎖音
        'ʲ': 'palatalization',   # 口蓋化
        'ɬ': 'lateral_fricative', # 側面摩擦音
        'χ': 'uvular_fricative',  # 口蓋垂摩擦音
        'x': 'velar_fricative',   # 軟口蓋摩擦音
        'q': 'uvular_stop',       # 口蓋垂閉鎖音
    }

    # 訓練データ統計（Phase 7-4 Fold 1）
    TRAINING_STATS = {
        'mean_duration': 3.0,      # 平均3秒
        'max_duration': 15.0,      # 最大15秒
        'mean_snr': 25.0,          # 平均SNR 25dB
        'min_snr': 15.0,           # 最小SNR 15dB
        'silence_ratio': 0.1,      # 無音割合10%
    }

    # CER予測モデル（経験的）
    CER_BASELINE = 5.23  # Fold 1のベースラインCER

    def __init__(self):
        pass

    def analyze_audio_quality(self, audio_data, sample_rate):
        """
        音声品質を分析

        Args:
            audio_data: 音声データ（numpy array）
            sample_rate: サンプリングレート

        Returns:
            dict: 分析結果
        """
        results = {}

        # 1. 音声の長さ
        duration = len(audio_data) / sample_rate
        results['duration'] = duration
        results['duration_warning'] = duration > self.TRAINING_STATS['max_duration']

        # 2. SNR（Signal-to-Noise Ratio）推定
        snr = self._estimate_snr(audio_data)
        results['snr'] = snr
        results['snr_warning'] = snr < self.TRAINING_STATS['min_snr']

        # 3. 音量レベル
        rms = np.sqrt(np.mean(audio_data**2))
        results['rms'] = rms
        results['volume_warning'] = rms < 0.01 or rms > 0.9

        # 4. 無音区間の割合
        silence_ratio = self._calculate_silence_ratio(audio_data)
        results['silence_ratio'] = silence_ratio
        results['silence_warning'] = silence_ratio > 0.3

        # 5. クリッピング検出
        clipping_ratio = np.sum(np.abs(audio_data) > 0.99) / len(audio_data)
        results['clipping_ratio'] = clipping_ratio
        results['clipping_warning'] = clipping_ratio > 0.01

        # 6. 周波数特性（スペクトル平坦度）
        spectral_flatness = self._calculate_spectral_flatness(audio_data, sample_rate)
        results['spectral_flatness'] = spectral_flatness

        # 7. 総合品質スコア（0-100）
        quality_score = self._calculate_quality_score(results)
        results['quality_score'] = quality_score

        return results

    def estimate_expected_cer(self, audio_quality, transcription=None):
        """
        期待CERを推定

        Args:
            audio_quality: analyze_audio_quality()の結果
            transcription: 転写結果（オプション、音素分析用）

        Returns:
            dict: 予測CER範囲と警告
        """
        base_cer = self.CER_BASELINE  # 5.23%

        # 品質スコアによるCER増加係数
        quality_score = audio_quality['quality_score']

        if quality_score >= 80:
            cer_multiplier = 1.0  # 5.23%
            confidence = "高"
        elif quality_score >= 60:
            cer_multiplier = 1.5  # 7.85%
            confidence = "中"
        elif quality_score >= 40:
            cer_multiplier = 2.5  # 13.08%
            confidence = "低"
        else:
            cer_multiplier = 4.0  # 20.92%
            confidence = "非常に低"

        # 音声長によるペナルティ
        duration = audio_quality['duration']
        if duration > 10:
            cer_multiplier *= 1.3  # 長い音声はエラー蓄積
        elif duration > 20:
            cer_multiplier *= 1.6

        # 困難音素によるペナルティ
        if transcription:
            difficult_ratio = self._calculate_difficult_phoneme_ratio(transcription)
            if difficult_ratio > 0.1:
                cer_multiplier *= (1 + difficult_ratio)

        predicted_cer = base_cer * cer_multiplier
        predicted_cer_range = (
            max(base_cer, predicted_cer * 0.7),
            min(50.0, predicted_cer * 1.3)
        )

        return {
            'predicted_cer': predicted_cer,
            'cer_range': predicted_cer_range,
            'confidence': confidence,
            'quality_score': quality_score,
            'factors': self._identify_cer_factors(audio_quality, transcription)
        }

    def _estimate_snr(self, audio_data):
        """SNRを推定（簡易版）"""
        # エネルギーベースの推定
        # 上位50%をシグナル、下位20%をノイズと仮定
        energies = audio_data ** 2
        signal_energy = np.percentile(energies, 75)
        noise_energy = np.percentile(energies, 25)

        if noise_energy > 0:
            snr = 10 * np.log10(signal_energy / noise_energy)
        else:
            snr = 60.0  # ノイズがほぼない

        return float(snr)

    def _calculate_silence_ratio(self, audio_data, threshold=0.02):
        """無音区間の割合を計算"""
        silence_samples = np.sum(np.abs(audio_data) < threshold)
        return silence_samples / len(audio_data)

    def _calculate_spectral_flatness(self, audio_data, sample_rate):
        """スペクトル平坦度を計算（ノイズ度の指標）"""
        try:
            spectral_flatness = librosa.feature.spectral_flatness(y=audio_data)[0]
            return float(np.mean(spectral_flatness))
        except:
            return 0.5  # デフォルト値

    def _calculate_quality_score(self, results):
        """総合品質スコアを計算（0-100）"""
        score = 100.0

        # SNRペナルティ
        if results['snr'] < 15:
            score -= 30
        elif results['snr'] < 20:
            score -= 15

        # 音量ペナルティ
        if results['volume_warning']:
            score -= 20

        # クリッピングペナルティ
        if results['clipping_warning']:
            score -= 25

        # 無音ペナルティ
        if results['silence_ratio'] > 0.3:
            score -= 15

        # 長さペナルティ
        if results['duration'] > 20:
            score -= 10
        elif results['duration'] > 30:
            score -= 20

        return max(0, score)

    def _calculate_difficult_phoneme_ratio(self, transcription):
        """困難音素の割合を計算"""
        if not transcription:
            return 0.0

        total_chars = len(transcription)
        if total_chars == 0:
            return 0.0

        difficult_count = sum(
            transcription.count(phoneme)
            for phoneme in self.DIFFICULT_PHONEMES.keys()
        )

        return difficult_count / total_chars

    def _identify_cer_factors(self, audio_quality, transcription):
        """CER増加要因を特定"""
        factors = []

        # 音声品質要因
        if audio_quality['snr_warning']:
            factors.append({
                'type': 'audio_quality',
                'name': 'ノイズが多い',
                'impact': '高',
                'description': f"SNR {audio_quality['snr']:.1f}dB（推奨: >15dB）"
            })

        if audio_quality['volume_warning']:
            factors.append({
                'type': 'audio_quality',
                'name': '音量が不適切',
                'impact': '中',
                'description': f"RMS {audio_quality['rms']:.3f}（推奨: 0.1-0.7）"
            })

        if audio_quality['clipping_warning']:
            factors.append({
                'type': 'audio_quality',
                'name': 'クリッピング検出',
                'impact': '高',
                'description': f"{audio_quality['clipping_ratio']*100:.1f}%のサンプルが飽和"
            })

        if audio_quality['silence_warning']:
            factors.append({
                'type': 'audio_quality',
                'name': '無音区間が多い',
                'impact': '低',
                'description': f"無音: {audio_quality['silence_ratio']*100:.1f}%"
            })

        if audio_quality['duration_warning']:
            factors.append({
                'type': 'duration',
                'name': '音声が長い',
                'impact': '中',
                'description': f"{audio_quality['duration']:.1f}秒（推奨: <15秒）"
            })

        # 音素要因
        if transcription:
            difficult_ratio = self._calculate_difficult_phoneme_ratio(transcription)
            if difficult_ratio > 0.1:
                factors.append({
                    'type': 'phoneme',
                    'name': '困難な音素が多い',
                    'impact': '高',
                    'description': f"困難音素: {difficult_ratio*100:.1f}%（ʔ, ʲ, ɬ, χなど）"
                })

        return factors

    def generate_quality_report_html(self, audio_quality, cer_prediction, actual_cer=None):
        """
        品質レポートのHTMLを生成

        Args:
            audio_quality: 音声品質分析結果
            cer_prediction: CER予測結果
            actual_cer: 実際のCER（オプション）

        Returns:
            str: HTML
        """
        quality_score = audio_quality['quality_score']

        # 品質スコアの色分け
        if quality_score >= 80:
            score_class = 'quality-excellent'
            score_emoji = '🟢'
        elif quality_score >= 60:
            score_class = 'quality-good'
            score_emoji = '🟡'
        elif quality_score >= 40:
            score_class = 'quality-fair'
            score_emoji = '🟠'
        else:
            score_class = 'quality-poor'
            score_emoji = '🔴'

        html = f'''
        <div class="quality-report">
            <div class="quality-header">
                <h3>📊 音声品質分析</h3>
                <div class="quality-score-box {score_class}">
                    {score_emoji} 品質スコア: <strong>{quality_score:.0f}/100</strong>
                </div>
            </div>

            <div class="quality-details">
                <div class="quality-item">
                    <span class="quality-label">SNR:</span>
                    <span class="quality-value">{audio_quality['snr']:.1f} dB</span>
                    {'⚠️ 低い' if audio_quality['snr_warning'] else '✓'}
                </div>
                <div class="quality-item">
                    <span class="quality-label">音声長:</span>
                    <span class="quality-value">{audio_quality['duration']:.1f} 秒</span>
                    {'⚠️ 長い' if audio_quality['duration_warning'] else '✓'}
                </div>
                <div class="quality-item">
                    <span class="quality-label">無音割合:</span>
                    <span class="quality-value">{audio_quality['silence_ratio']*100:.1f}%</span>
                    {'⚠️ 多い' if audio_quality['silence_warning'] else '✓'}
                </div>
                <div class="quality-item">
                    <span class="quality-label">クリッピング:</span>
                    <span class="quality-value">{audio_quality['clipping_ratio']*100:.2f}%</span>
                    {'⚠️ 検出' if audio_quality['clipping_warning'] else '✓'}
                </div>
            </div>

            <div class="cer-prediction">
                <h4>🎯 予測CER範囲</h4>
                <div class="predicted-cer">
                    <strong>{cer_prediction['cer_range'][0]:.1f}% - {cer_prediction['cer_range'][1]:.1f}%</strong>
                    （中央値: {cer_prediction['predicted_cer']:.1f}%）
                </div>
                <div class="confidence">
                    信頼度: <strong>{cer_prediction['confidence']}</strong>
                </div>
        '''

        # 実際のCERとの比較
        if actual_cer is not None:
            actual_cer_pct = actual_cer * 100
            in_range = cer_prediction['cer_range'][0] <= actual_cer_pct <= cer_prediction['cer_range'][1]

            html += f'''
                <div class="actual-cer-comparison">
                    <div class="actual-cer">
                        実際のCER: <strong>{actual_cer_pct:.2f}%</strong>
                        {' ✓ 予測範囲内' if in_range else ' ⚠️ 予測範囲外'}
                    </div>
            '''

            if not in_range:
                if actual_cer_pct < cer_prediction['cer_range'][0]:
                    html += '<div class="analysis-note">📈 予測より良好！この音声は訓練データに近い可能性があります。</div>'
                else:
                    html += '<div class="analysis-note">📉 予測より悪化。下記の要因を確認してください。</div>'

            html += '</div>'

        html += '</div>'  # cer-prediction

        # CER増加要因
        if cer_prediction['factors']:
            html += '<div class="cer-factors"><h4>⚠️ CER増加要因</h4><ul>'
            for factor in cer_prediction['factors']:
                impact_emoji = {'高': '🔴', '中': '🟠', '低': '🟡'}.get(factor['impact'], '⚪')
                html += f'''
                    <li class="factor-item">
                        <span class="factor-name">{impact_emoji} {factor['name']}</span>
                        <span class="factor-impact">影響度: {factor['impact']}</span>
                        <div class="factor-description">{factor['description']}</div>
                    </li>
                '''
            html += '</ul></div>'

        # 推奨事項
        html += '<div class="recommendations"><h4>💡 推奨事項</h4><ul>'

        if audio_quality['snr_warning']:
            html += '<li>ノイズ除去処理を試してください</li>'
        if audio_quality['duration_warning']:
            html += '<li>音声を短い区間（<15秒）に分割することを検討してください</li>'
        if audio_quality['volume_warning']:
            html += '<li>音声の音量を調整してください（正規化）</li>'
        if audio_quality['clipping_warning']:
            html += '<li>クリッピングのない音源を使用してください</li>'

        if not cer_prediction['factors']:
            html += '<li>✓ 音声品質は良好です！</li>'

        html += '</ul></div>'
        html += '</div>'  # quality-report

        return html