Note:
In the speaker id community the words “train”, “test” and “development”
are used in a different sense from in the speech recognition community. In
speaker-id land, the “development” data is the data you actually use to build
the system; the “training” and “test” data are really both used to test how
well the systme performs. The “training” data is used as enrolment data for a
particular speaker, and the evaluation consists of taking utterances from the
“test” data and having the system decide whether they were uttered by various
training speakers.(来自于kaldi/egs/sre08/v1/README.txt)
短语音对于i-vector的训练是有害的,小于100k的utterence都会被删除。(来自于kaldi/egs/sitw/v1)
enroll集和eval集可能有下面不同:
- 跨信道,跨设备,跨距离(ios录的?安卓录的?近场录的?远场录的?1个mic录的?n个mic录的?)
- 性别比例分布极不平衡
- 采样率一样么?(16k?44.1K?)
- 音量是否有过大过小情况?
- 语速是否有过快过慢情况?
- 时长差异大么?
- 是否有情绪变化包含在内?
- 是否有仿冒攻击情况?