Multilingual Speech Processing

44 CHAPTER 3. LINGUISTIC DATA RESOURCES

3.3 Data Collection Efforts in the United States

The past decade has seen significant effort devoted toward multilin-

gual speech processing in the United States. Much of this effort has

been associated with annual metrics-based technology evaluation projects

administered by the National Institute for Standards and Technology

(NIST) and sponsored by the Defense Advanced Research Projects

Agency (DARPA). In the following sections some of the larger efforts—

including the CallHome, CallFriend, Switchboard, Fisher, and Mixer

collections—will be discussed.

Space limitations prevent us from giving adequate treatment to the

efforts of several groups that have created important linguistics resources.

The Center for Spoken Language Understanding (CSLU) at the Oregon

Graduate Institute of Science and Technology has published 20 different

corpora since 1992. Two are collections of telephone speech involving mul-

tiple languages: EasternArabic, Cantonese, Czech, English, Farsi, French,

German, Hindi, Hungarian, Japanese, Korean, Malay, Mandarin, Italian,

Polish, Portuguese, Russian, Spanish, Swedish, Swahili, Tamil and Viet-

namese. The Johns Hopkins Center for Language and Speech Processing

(CLSP) typically develops one or more databases each year through its

annual summer workshops. The Institute for Signal and Information Pro-

cessing (ISIP) of the Mississippi State University has contributed several

important language resources, including software for echo cancellation,

a speech recognition tool kit, a corpus of southern accented speech, the

resegmentation of the Switchboard corpus, and JEIDA—a collection of

prompted Japanese speech. ISIP resources are available from its home

page or the Linguistic Data Consortium. Last but not least, the U.S. Military

Academy’s Center for Technology Enhanced Language Learning (CTELL)

has created speech corpora in Arabic, Russian, Portuguese, American

English, Spanish, Croatian, Korean, German, and French along the way to

creating recognition systems to support language learning at West Point.

The Linguistic Data Consortium addresses the needs of researchers in

speech and written language processing by licensing, collecting, creating,

annotating, and sharing linguistic resources, including data, tools, stan-

dards, and best practices. Since its creation in 1992, the LDC has distributed

more than 25,000 copies of more than 300 titles and otherwise shared data

with 1,820 organizations in 93 countries. The LDC often serves as data

coordinator for NIST technology evaluation and DARPA common task

3.3. DATA COLLECTION EFFORTS IN THE UNITED STATES 45

programs. This role is not assigned by fiat but is decided on the basis of com-

petition. Although the LDC is located within the United States, its mem-

bership is open to researchers around the world. More than half of all LDC

shipments have had destinations outside the United States. The LDC cat-

alog lists all of the corpora LDC has released and may be searched by the

languages of the corpus, types and sources of data included, recommended

uses, research programs for which it is relevant, and release year.

6

3.3.1 CallHome

The CallHome collections supported the large-vocabulary conversational

speech recognition (LVCSR) program (NIST, 1997) in which researchers

built systems to automatically transcribe large vocabulary, continuous

speech, specifically conversational telephone speech. They represent

perhaps the earliest effort to define and create a basic resource kit

for developers of a specific linguistic technology. For each CallHome

language—English, Mandarin Chinese, Egyptian Colloquial Arabic,

Spanish, Japanese, and German—at least 200 international telephone

conversations, 20 to 30 minutes in duration, were collected. Subjects par-

ticipated in a single call, speaking to partners about topics of their choosing.

All calls originated within and terminated outside of the United States.

Parts of each of the calls in each language were transcribed orthograph-

ically and verbatim. If a call was designated as training or development test

data, 10 minutes were transcribed. If a call was designated as evaluation

data, only 5 minutes were transcribed. For English, Spanish, and German,

verbatim, orthographic transcription was a straightforward, if challeng-

ing, task. However, Mandarin Chinese, Egyptian Colloquial Arabic, and

Japanese presented additional challenges. Within the Japanese transcripts,

word segmentation was performed by hand. The Mandarin Chinese tran-

scripts were automatically segmented using software developed at the

Linguistic Data Consortium. Egyptian Colloquial Arabic offered perhaps

the largest challenge since it is not generally written. Researchers at the

Linguistic Data Consortium needed to invent a writing system for this

variety and then train native speakers to use it consistently before tran-

scription could begin. The verbatim transcripts include every word and

6

See http://www.ldc.upenn.edu/catalog