CSR CORPUS COLLECTION 
Denise Danielson, Project Leader 
Jared Bernstein, Principal Investigator 
SRI International 
Menlo Park, California 94025 
PROJECT GOALS 
The objective of the CSR Corpus Development is to collect 
and deliver a large corpus of continuous speech data to sup- 
port DARPA research efforts in continuous speech recogni- 
tion (CSR). SRI's current goal is the completion of Phase 2, 
Part 1 of the planned CSR Corpus. This consists of 86,000 
sentences from 275 speakers, including 8000 spontaneous 
sentences from 40 joumatists. 
The Phase 2 Corpus collection task is a high volume data 
production task. SRI's major goal has been efficiency. 
Other goals include gathering data that is more representa- 
tive of the real world by minimizing controls on vocabu- 
lary, microphones, background noise and speaker 
disfluencies, while improving data quality controls. 
RECENT RESULTS 
SRI began work on the current phase of CSR in September, 
1992 and expects to complete delivery of this portion in 
June, 1993. 
Data Production -- As of 12 March 1993, SRI has col- 
lected the following portion of this CSR database: 
Subject type Required Collected 
Non-journalist long 25 22 
Non-journalist short 210 125 
Journalist 40 17 
Subject Efficiency -- SRI's first goal was to speed up sub- 
ject interaction with the data collection software. Addi- 
tional memory was added to the data collection systems, 
and data collection software made much faster, so that now 
the pace of the data collection process is directly controlled 
by the subject and no longer limited by the software. As a 
result, the average data collection pace has increased from 
125 utts/hr to 200 utts/hr. For a typical short-term non-jour- 
nalist subject collecting 190 read sentences, these changes 
and a faster paced orientation have reduced subject time 
from 120 minutes to 90 minutes. The shorter time require- 
ment also makes it easier to attract and schedule subjects. 
Process Efficiency -- SRI has also been concerned with 
reducing the labor required to process speech data. A labor 
savings was realized by removing monitors from the data 
collection room. The data collection monitor now spends 
about 25 minutes instructing and observing while subjects 
collect their first few utterances, and then leaves the room. 
Two other changes have significanly improved labor effi- 
ciency. SRI has developed a new transcription tool that has 
led to a 15% to 20% reduction in transcription time and 
improved accuracy. We have also automated most pre- 
archival and archival steps. 
Data Quality -- SRI has incorporated NIST data quality 
software into its procedures. Sample files are collected at 
the start of each day on each data collection system. These 
files are run through the wavmd program, which runs a sig- 
nal-to-noise (SNR) evaluation and other tests. Additional 
checks are performed on all files as they are collected to 
ensure that problems (e.g. dead microphone) are caught. 
Labor Analysis -- SRI is analyzing labor costs as we pro- 
ceed with the current project to enable us to predict costs in 
the future, as well as to target specific tasks for efficiency 
improvements. A first round of labor analysis in January of 
this year identified transcriptioa as one of the biggest labor 
costs. This has led to efforts to make the transcription task 
easier and more efficient. SRI continues to work with NIST 
and the CCCC to clarify transcription guidelines and 
implement changes recommended by CCCC. An analysis 
of recent project labor indicates that 10-15% of SRI's CSR 
project time has been spent on tasks in support of commu- 
nication with NIST and various DARPA program commit- 
tees. 
PLANS FOR THE COMING YEAR 
• Collect the remainder of the CSR Corpus. 
• Work with NIST and CCCC to define goals and con- 
straints for alternate microphone, s and environments. 
• Work with NIST and the Data Quality Committee to fur- 
thex improve and automate quality tests of speech files. 
• Work with NIST and the Data Quality Committee to 
define documentation requirements. 
• Refine the spontaneous speech collection paradigm. 
413 
