# 示意代码,不可直接运行:仅演示韵律/声学特征提取的分类思路,并非论文官方实现。
# 论文实际方法为 Whisper ASR 嵌入 + LLM 语言特征融合(见 arXiv:2606.30675 / 2606.28445),
# 下面的特征分类与认知评分阈值为本文撰写,未经验证。
import numpy as np

class VoiceBiomarkerExtractor:
    """
    语音生物标记物提取器(示意)

    展示面向认知状态评估的韵律/音质/语言/情感等特征分类思路
    """
    
    FEATURE_CATEGORIES = {
        '韵律特征': {
            '语速': 'words_per_minute, syllables_per_second',
            '停顿': 'pause_count, pause_duration_mean, pause_ratio',
            '音调': 'pitch_mean, pitch_range, pitch_slope',
            '能量': 'energy_mean, energy_range, shimmer',
        },
        '音质特征': {
            '嗓音': 'jitter, shimmer, HNR',
            '呼吸': 'breath_count, breath_pattern',
            '紧张度': 'spectral_tilt, formant_frequency',
        },
        '语言特征': {
            '词汇': 'vocabulary_richness, TTR, MTLD',
            '语义': 'semantic_similarity, content_density',
            '流畅性': 'disfluency_count, filler_words',
            '语法': 'syntax_complexity, sentence_length',
        },
        '情感特征': {
            '情感效价': 'valence_mean, valence_range',
            '唤醒度': 'arousal_mean, arousal_range',
            '情感表达': 'emotional_diversity',
        },
        '时序特征': {
            '节奏': 'rhythm_regularity, speech_tempo',
            '组织性': 'narrative_structure, coherence',
        },
        '声学特征': {
            'MFCC': '13-dim MFCC + delta + delta-delta',
            '频谱': 'spectral_centroid, spectral_flux',
            '共振峰': 'F1, F2, F3 frequencies',
        },
        '内容特征': {
            '转录': 'ASR transcript → BERT embedding',
            '关键词': 'keyword_density, topic_relevance',
        }
    }
    
    def __init__(self, fs: int = 16000):
        self.fs = fs
    
    def extract_prosody(self, audio: np.ndarray) -> dict:
        """提取韵律特征"""
        # 基频提取
        from scipy.signal import medfilt
        
        # 简化的基频估计（自相关法）
        def estimate_f0(signal, fs, frame_len=0.04, hop=0.02):
            frame_samples = int(fs * frame_len)
            hop_samples = int(fs * hop)
            f0_list = []
            
            for i in range(0, len(signal) - frame_samples, hop_samples):
                frame = signal[i:i+frame_samples]
                if np.max(np.abs(frame)) < 0.01:
                    f0_list.append(0)
                    continue
                
                # 自相关
                autocorr = np.correlate(frame, frame, mode='full')
                autocorr = autocorr[len(autocorr)//2:]
                
                # 找第一个峰（跳过零延迟峰）
                min_lag = int(fs / 400)  # 400Hz max
                max_lag = int(fs / 80)   # 80Hz min
                
                if max_lag >= len(autocorr):
                    f0_list.append(0)
                    continue
                
                peak_region = autocorr[min_lag:max_lag]
                if len(peak_region) == 0 or np.max(peak_region) < 0.1:
                    f0_list.append(0)
                    continue
                
                peak_idx = np.argmax(peak_region) + min_lag
                f0 = fs / peak_idx
                f0_list.append(f0)
            
            return np.array(f0_list)
        
        f0 = estimate_f0(audio, self.fs)
        f0_voiced = f0[f0 > 0]
        
        # 停顿检测（能量阈值）
        frame_energy = np.array([
            np.sum(audio[i:i+int(self.fs*0.02)]**2)
            for i in range(0, len(audio)-int(self.fs*0.02), int(self.fs*0.01))
        ])
        threshold = 0.1 * np.max(frame_energy)
        pauses = frame_energy < threshold
        pause_ratio = np.sum(pauses) / len(pauses)
        
        return {
            'pitch_mean': float(np.mean(f0_voiced)) if len(f0_voiced) > 0 else 0,
            'pitch_range': float(np.std(f0_voiced)) if len(f0_voiced) > 0 else 0,
            'pause_ratio': float(pause_ratio),
            'pause_count': int(np.sum(np.diff(pauses.astype(int)) == 1)),
            'speech_rate': float(1 - pause_ratio),  # 简化
        }
    
    def cognitive_score(self, features: dict) -> float:
        """
        综合认知评分 (0-100)
        0 = 严重认知损伤
        100 = 认知健康
        """
        score = 50.0
        
        # 语速：正常 130-180 wpm
        rate = features.get('speech_rate', 0.5)
        if rate < 0.3:
            score -= 15  # 语速过慢
        elif rate > 0.8:
            score += 5
        
        # 停顿比例：正常 20-35%
        pause = features.get('pause_ratio', 0.5)
        if pause > 0.5:
            score -= 20  # 停顿过多
        elif pause < 0.2:
            score += 5
        
        # 音调范围：单调→可能损伤
        pitch_range = features.get('pitch_range', 0)
        if pitch_range < 10:
            score -= 10  # 单调
        
        return float(np.clip(score, 0, 100))


# 测试
if __name__ == "__main__":
    rng = np.random.default_rng(42)
    extractor = VoiceBiomarkerExtractor(fs=16000)
    
    # 模拟正常语音
    t = np.arange(0, 5, 1/16000)
    normal_speech = 0.3 * np.sin(2*np.pi*150*t) * (np.sin(2*np.pi*3*t) > 0.5)
    normal_speech += 0.05 * rng.normal(size=len(t))
    
    # 模拟疲劳/认知受损语音（更慢、更多停顿、更单调）
    tired_speech = np.zeros_like(t)
    for i in range(0, len(t), 16000):  # 每秒一个短段
        seg_end = min(i + 8000, len(t))  # 0.5s说话
        tired_speech[i:seg_end] = 0.2 * np.sin(2*np.pi*120*t[i:seg_end])
    
    normal_feat = extractor.extract_prosody(normal_speech)
    tired_feat = extractor.extract_prosody(tired_speech)
    
    print("=== 正常语音特征 ===")
    for k, v in normal_feat.items():
        print(f"  {k}: {v:.3f}")
    
    print("\n=== 疲劳语音特征 ===")
    for k, v in tired_feat.items():
        print(f"  {k}: {v:.3f}")
    
    print(f"\n正常认知评分: {extractor.cognitive_score(normal_feat):.1f}")
    print(f"疲劳认知评分: {extractor.cognitive_score(tired_feat):.1f}")