Rohan Paul· @rohanpaul_ai · X·· 3 小时前AI 评分29
AI 导读
更有用的测试应该混合地区口音、非母语者、背景噪声、远场录音、快速语速、犹豫、自我纠正、领域词汇、手机麦克风和语码转换。 然后把基准真值发布在原始转写旁边,检查实际的替换、删除、插入和幻觉。
正文
A more useful test would mix regional accents, non-native speakers, background noise, far-field recordings, fast speech, hesitations, self-corrections, domain vocabulary, phone microphones, and code-switching.
Then publish the ground truth beside the raw transcript and inspect the actual substitutions, deletions, insertions, and hallucinations.
来源:Rohan Paul · x.com