毕设-周报-20150520

上周：
基线的ROC曲线去要重新确定判定标准，发往论文作者处邮件回复如下（版权为论文作者所有）：

The curve itself doesn’t matter. I was even suggested by the reviewers to use DET curves as they can better show the results.

For detections, we work on the sentence level. For example, if your sentence contains the keyword, and your decision for detection is YES, then you get a correct detection. We didn’t check the alignment explicitly (i.e., if the boundary of the detected keyword matches the keyword in the sentence exactly), but we only have short sentences, so I guess that’s OK. We make one YES/NO decision for each sentence.

False alarm rate = # of false alarms / # of sentences
False rejection rate = # of false rejections / # of sentences (this is different from the traditional ROC curve)

对下一步大词汇量训练的建议如下：

es I’d suggest to use Librispeech, as we have 1000 hours for that. You’ll have to do forced alignment to generate those time information. If you train your system with Librispeech, I’m sure you’ll have training data alignment somewhere, and you can use that. For evaluation data, you can use your trained model and the reference to do the alignment, and that should be sufficient I think (and that’s what we did for the paper, but we worked on another dataset that is not publicly available).

上次周会至今在进行Librispeech test 训练集上的切词工作(切出模版)，尽量选取原作者使用的若干关键词进行实验，同时修改绘制评测曲线的PYTHON脚本。结果整理在下篇推出。

毕设结尾，在尝试使用android转移基线系统的音素识别器做demo，同时继续大词汇量下的训练。