送交者: articsyn 于 2006-1-31, 03:51:08:
My friend told me there are some people discussing
speech recognition here. I just took a look and
want to say something. I found it was wasting my
time because my interest is research but it seems
more individual attack here. So this is my first
and last post.
I've been working in speech recognition for several
years. Now I'm focusing on other aspects of speech
processing. If anyone here has reviewed the literature
on speech recognition from 1970's, you'll find the "95%"
accuracy rate could often be achieved UNDER CERTAIN
TEST CONDITIONS. To my understanding, the accuracy rate
range is highly dependent on the task, and only
comparable among the systems that joined the test.
Here is an example link in "《新清华》2002年9月27日":
http://news.cic.tsinghua.edu.cn/xqh/xqhnews/read.php?id=4170
Please notice the numbers "93.7%" and "98.7%".
If you don't understand the test conditions, the
numbers are nonsense to you.
I had studied in Tsinghua and the Institute of
Automation, and worked in IBM's speech lab for a period
of time. My previous colleagues and I had never looked
down upon any researchers in other local teams. To those
who blindly pre-assumed the technique of speech recognition in a local institute SHOULD be worse
than that of oversea labs, I really disrespect them :(
Maybe people dislike the news style in CCTV (me too).
I disagree with a talk-show that presents a single
statistical number to the general public without any detailed explanation either. But it's useless to suspect
everything and do nothing constructive except for
individual attacking. It's even harmful to conclude without any background research and knowledge. If anyone here really care about a scientific evaluation, please go to study and compare the 863-evaluation procedure with the NIST one. If you find something can be improved, that's a real academic contribution!