比起默认的turbo有很大改观,不过也有新的问题,转录文本的前/后有莫名的padding

#9
by puppyinwind - opened

用pipeline处理一段莫言的发言片段,结果比起原生的turbo改良太多(准确度和标点都很不错)- 感谢作者 :)
但新问题也出来了,可以看到转录文字的开始和结束处,出现了无意义的字母(如“DiK,O,O.M,O,”),无意义的短语(如你。俩。你。。。你)和标点符号(最后一大堆的句号):
DiK,O,O.M,O,已经有了或多或少的了解,你们。也许看到了我九十岁的老父亲,看到我的哥哥姐姐,我的妻子女儿和我一岁零四个月的外甥。,当然有一个我此刻最想念的人我的母亲。你。俩。你。。。你。。。。。。。。。。。。,。。。。。。。。。。。。。。。。,。。。。。。。。。。K。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。。
原始的mp3附上:

又用其他几个长短不一的音频测试,发现都有这个问题,幻觉混乱比较典型地集中在文本的开端和末尾,比如一个央视新闻音频的结尾,自动添加了这段“。d嘛的地方啊在你中。的风向。啊。.啊。.好看多久也不见。啊。对夜光。爱。.meM..D.L.A.D.C.A.N.M.Y.J.D.L.A.D.O.M.M.M.M.M.L.M.M.M.M.M.M.M.M.M.M.M.M.M.M.L.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M.M。M.M.M.M.M.M.M.M.M.M.M.M.M.MM。M. 哦,不。好啊。.北岛。”

能否帮忙看看是怎么回事?感谢感谢

同样出现该问题,原始mp3

结果

三D的D,目前看呃,从最原始D,单层工艺到多层累积,但是它的一个呃引脚还是走平面。那
么这一个只是单纯或者看那个单色体带,它的一个是没有变的,因此物质取给它布线的这么一个面积,那它布线还是周围的一边。那再到这个叫HMC吧。 Hi, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then, and then.那么这一个相当于是啊充分利益,将这个速度的 扩充到了这么一个面积。但是实际上呢它相对于这个堆叠来说,这么多个堆叠层呢,实际上还是复用了这么一个是会接口。那么更进一步,那我就是去做呃我要去做福才一马瑞化。但实际上福才是一马瑞很难去做到外部照,也没有去在堆叠层上去做。那么可能更多更是在底层去做。但是如果如果在底层去做的话,实际上你在底层做一个老师可大嗯,然后跟上面这样拼一起的话啊,那这种操作实际上可能没有太利用到这么一个啊,片内的一个带宽。因为它变成带宽,实际上在三三立上雷起来了。那么我这一块其实没太高度,他做普什尔远瑞去用了这个带去做,那肯要去读一下那国的文章。另外一个就是呃如果像是这种M,它里面它不需要改变太多。我看它和M四它其实是兼容的。那么它其那这样的话实际上是每一层去做一 块,那这个也许它能够提到更好的一个效益。

Sign up or log in to comment