测评接口评分不准确

[mywsat]

1. 我使用仓库中的train 最终的训练结果是
   训练 Epoch 36/40: 100%|########| 40/40 [00:10<00:00, 3.78it/s, loss=1.7837, acc=45.9%, batch=40/40]
   训练损失: 1.7121
   训练准确率: 45.91%
   验证 Epoch 36/40: 100%|#############################################| 40/40 [00:09<00:00, 4.28it/s]
   验证损失: 1.9522
   验证准确率: 39.94%

   难音素准确率:
   L : 18.4% (3428/18658)
   R : 23.9% (3875/16223)
   OW : 31.3% (4741/15167)
   AW : 11.9% (843/7076)
   TH : 0.1% (8/5615)
   DH : 29.8% (2794/9365)
   V : 15.1% (1130/7492)
   W : 60.2% (8848/14700)
   ⚠️ 验证损失未改善 (5/5)
   要达到多少的准度才能达到算法的要求呢

2. 调用/api/pronunciation/assess 的评估
   音频文件是通过 Kokoro 82m生成的 评分的效果太差了。对此有什么改进的建议，尤其是针对多读 和漏读的判断对齐是否有别的方案
   {
   "overallScore": 47.1,
   "accuracyScore": 45.8,
   "fluencyScore": 64,
   "completenessScore": 25,
   "duration": 1.1,
   "wordCount": 1,
   "phonemeCount": 4,
   "words": [
   {
   "word": "hello",
   "score": 45.8,
   "confidence": 0.4,
   "startTime": 0.3,
   "endTime": 1.1,
   "duration": 0.8,
   "errorType": "Mispronunciation",
   "phonemes": [
   {
   "phoneme": "HH",
   "score": 88.1,
   "confidence": 0.8,
   "startTime": 0.3,
   "endTime": 0.5,
   "duration": 0.2,
   "gopScore": 1.4,
   "targetProb": 0.6,
   "confusionProb": 0.1,
   "errorType": "None",
   "nbestPhonemes": [
   {
   "phoneme": "HH",
   "score": 100
   },
   {
   "phoneme": "T",
   "score": 35.4
   },
   {
   "phoneme": "K",
   "score": 23.4
   },
   {
   "phoneme": "P",
   "score": 19.3
   },
   {
   "phoneme": "AE",
   "score": 7.8
   }
   ]
   },
   {
   "phoneme": "AH0",
   "score": 38.2,
   "confidence": 0.3,
   "startTime": 0.5,
   "endTime": 0.5,
   "duration": 0.1,
   "gopScore": -0.7,
   "targetProb": 0.2,
   "confusionProb": 0.4,
   "errorType": "Mispronunciation",
   "nbestPhonemes": [
   {
   "phoneme": "AW",
   "score": 89.9
   },
   {
   "phoneme": "AH",
   "score": 45.2
   },
   {
   "phoneme": "AA",
   "score": 31.9
   },
   {
   "phoneme": "AE",
   "score": 18.7
   },
   {
   "phoneme": "OW",
   "score": 14.6
   }
   ]
   },
   {
   "phoneme": "L",
   "score": 17.1,
   "confidence": 0.2,
   "startTime": 0.5,
   "endTime": 0.7,
   "duration": 0.2,
   "gopScore": -1.7,
   "targetProb": 0.1,
   "confusionProb": 0.4,
   "errorType": "Mispronunciation",
   "nbestPhonemes": [
   {
   "phoneme": "W",
   "score": 100
   },
   {
   "phoneme": "OW",
   "score": 59.3
   },
   {
   "phoneme": "L",
   "score": 19.5
   },
   {
   "phoneme": "AH",
   "score": 11.3
   },
   {
   "phoneme": "N",
   "score": 10.4
   }
   ]
   },
   {
   "phoneme": "OW1",
   "score": 39.7,
   "confidence": 0.4,
   "startTime": 0.7,
   "endTime": 1.1,
   "duration": 0.4,
   "gopScore": -0.4,
   "targetProb": 0.2,
   "confusionProb": 0.2,
   "errorType": "Mispronunciation",
   "nbestPhonemes": [
   {
   "phoneme": "AH",
   "score": 61.4
   },
   {
   "phoneme": "OW",
   "score": 41.4
   },
   {
   "phoneme": "L",
   "score": 27.1
   },
   {
   "phoneme": "ER",
   "score": 13.6
   },
   {
   "phoneme": "AA",
   "score": 12.4
   }
   ]
   }
   ]
   }
   ]
   }

3. 对齐实际过于耗时

4. 16:16:04,549 - 开始MFA对齐
   16:17:22,431 - MFA对齐执行成功（耗时：77.88秒）
   16:17:22,442 - 开始音素评分
   16:17:22,473 - 评分完成（耗时：0.03秒）
   总耗时：79.42秒
   最后感谢的分析 提供了我一些思路