✅ [1] 기본 점수 구조

score_dict = {
    "pitch_dtw": 72.68,
    "transpose": 70.0,
    "pitch_chunk": 75.62,
    "mfcc_cos": 0.0,
    "mfcc_norm_euc": 97.64,
    "mfcc_dtw": 33.39,
    "onset_euc": 99.02,
    "onset_dtw": 74.02,
    "onset_proxy": 63.7,
    "onset_match": 67.66,
    "tempo_abs": 27.51,
    "tempo_ratio": 8.83,
    "tempo_log": 18.88,
    "tempo_thresh": 60.0
}

✅ [2] 대표적인 합성 점수 계산식들

🎯 1. 단순 평균

total = sum(score_dict.values()) / len(score_dict)

🎯 2. 가중 평균

weights = {
    "pitch_dtw": 0.12, "transpose": 0.12, "pitch_chunk": 0.06,
    "mfcc_cos": 0.10, "mfcc_norm_euc": 0.10, "mfcc_dtw": 0.13,
    "onset_euc": 0.066, "onset_dtw": 0.066, "onset_proxy": 0.066,
    "onset_match": 0.066, "tempo_log": 0.066
}
weighted_sum = sum(score_dict[k] * weights.get(k, 0) for k in score_dict)
total_weight = sum(weights.get(k, 0) for k in score_dict)
final_score = round(weighted_sum / total_weight, 2)

🔍 지금까지는 이걸 쓰고 있었지!


🎯 3. Top-K 평균

k = 10
top_k = sorted(score_dict.values(), reverse=True)[:k]
top_k_avg = sum(top_k) / len(top_k)

🎯 4. Z-score 기반 정규화 후 평균

import numpy as np

values = np.array(list(score_dict.values()))
z_scores = (values - values.mean()) / values.std()
adjusted = 100 / (1 + np.abs(z_scores))  # z가 클수록 점수 낮아짐
z_score_avg = adjusted.mean()

🎯 5. 비선형 함수 (ex: 제곱 or 로그 강조)

nonlinear_score = sum([v ** 1.3 for v in score_dict.values()]) / len(score_dict)