英语翻译3.Speech recognition testingThe accuracy of several SR s

英语翻译
3.Speech recognition testing
The accuracy of several SR systems has been shown
in the literature.Some systems have been tested in a
constrained domain (benefiting from a restricted voca-
bulary) and have achieved error rates of less than 20%
[13,27],while other tests of commercial systems have
been anecdotal and tuned by the software vendors to
perform well on the particular test.Further,results
from tests of SR programs can vary,based upon the
complexity of the input[1].
3.1.Tests of multiple systems
Tests have shown that the accuracy for PC-based SR
systems typically ranges from 80 to 99%[2,4].While
some of these tests are not directly comparable,
because of differences in source text,efficiency and
effectiveness measures,and testing periods,they give
an indication of the current state of the technology.
Using standardized protocols and reference test
sets developed by the National Institute of Standards
[26],a mixture of research and commercially available
systems were compared directly,as shown inTable 2.
There was no statistically significant difference in the
results of the first three systems and little difference in
the first seven.However,the last five systems were
clearly less accurate.
3.2.Tests of DragonSystems
DragonSystems is perhaps the US market leader of
PC-based,continuous-speech,general-purpose SR
software.In a test at PC Labs with five subjects using
the software,an accuracy of 89.1% was achieved at 74
wpm as compared to 87.8% at 69 wpm with SR
software from IBM.2In addition,anecdotal tests of
DragonSystems have shown:3
At Raytheon,the average time to process a Material
Requisition Form was reduced from 162 to 44 s
(over 4 h per month) with the aid of voice macros.
An attorney’s staff increased their document pre-
paration productivity by 25%.
An attorney generated text at 120 wpm with 95%
accuracy using the software,and a nurse generated
138 wpm as compared to 40 wpm when typing.
Medical transcripts were produced in one-third of
the time required when typing.
A doctor reduced turnaround time for medical
transcriptions from 2 weeks to 1 day and reduced
his variable transcript service cost from US$ 150
per day to almost zero.
Total medical transcription charges of about US$
100,000 per year at a medical center were reduced
by about 80%.
DragonSystems’NaturallySpeaking Preferred,Ver-
sion 4.01 was selected for our study.Other,perhaps
more accurate,commercial SR systems,such as
LVCR,LIMSI,and CU-HTK,are not readily obtain-
able,and as a market leader,use of the selected system
makes the study more replicable.
呆头鹰 1年前 已收到1个回答 举报

哇来了 幼苗

共回答了17个问题采纳率:94.1% 举报

3 语音识别测试
文献中记载了几个语音识别系统的精确度.一些系统是基于特定的域(偏向于某一类特定词汇)进行测试的,错误率已低于20%.而其他一些商用系统则是通过软件的方式人为的提高在某一类测试中的正确率.而且,输入测试样本的复杂度不同也会使测试结果差别很大.
3.1 多系统测试
测试显示,基于PC机的语音识别系统识别率大致在80%-99%之间.而这其中的有些测试并不是直接可比较的,因为输入的样本不同和度量的标准都不同.而测试的时间也有影响,通常测试只反映出在当时的技术条件下所得到的结果.利用标准的协议和国家标准化组织给出的测试集,我们对比了供研究的系统和商用系统的结果,如表2所示.从统计学上讲,前三个系统的结果基本没有差别,而前七个系统的结果只有微小差别.然而,后5个系统的结果明显更加不精确.
有急事,剩下的回头再翻译

1年前

6
可能相似的问题
Copyright © 2024 YULUCN.COM - 雨露学习互助 - 16 q. 0.023 s. - webmaster@yulucn.com