EARS

The DARPA EARS (Effective, Affordable, Reusable Speech-to-Text) program is developing robust speech recognition technology to address a range of languages and speaking styles and "produce powerful new speech-to-text (automatic transcription) technology whose outputs are substantially richer and much more accurate than currently possible." LDC provides conversational and broadcast speech and transcripts, annotations, and lexicons and texts for language modeling in each of the EARS languages.

EARS Data Matrix – Revised 11/07/04

EARS Arabic Website –

RT-04 Data and Annotation Plan – Revised 07/15/04

Quick Transcription
LDC is using a quick transcription approach in order to provide large volumes of transcribed speech for training purposes. Quick transcription limits the amount of time an annotator spends transcribing each file, and removes some level of detail from the resulting transcript. The following documents provide additional information about quick transcription for STT research:

RT-03 Q uick transcription guidelines

Careful Transcription
LDC creates high-accuracy verbatim transcripts with rich markup for use as Speech-to-Text devtest and evaluation data. The following documents provide additional information about careful transcription for STT research:

RT-04 Careful Transcription Guidelines in .pdf or .doc format

Metadata Extraction
LDC is creating annotated resources to support a metadata extraction (MDE) research evaluation. The goal of MDE is to enable technology that can take the raw STT output and refine it into forms that are of more use to humans and to downstream automatic processes. The following documents provide additional information about annotation in support of MDE research:

MDE Annotation Project Website
SimpleMDE Annotation Guidelines V6.2 in .pdf format

The table below summarizes English, Chinese and Arabic resources relevant to EARS categorized by language and resource type.

	English
Broadcast Speech	1997 HUB4 English Evaluation Speech and Transcripts - 3 hours of radio and tv news stories 1996 English Broadcast News Speech (Hub-4) - 104 hours of broadcasts 1996 English Broadcast News Dev and Eval - 104 hours of broadcasts 1997 English Broadcast News Speech (Hub-4) - 97 hours of news broadcasts TDT2 English Audio - 1036 waveform files of complete broadcasts TDT3 English Audio - 950 hours of news broadcasts TDT2 Careful Transcription Audio - recorded broadcasts from 1998 EARS-only training data releases (contact LDC) LDC2003E02 TDT4 Multilanguage Speech LDC2003E02A TDT4 Multilanguage Supplemental
Careful Transcriptions	1996 English Broadcast News Transcripts (Hub-4) - 104 hours of transcripted broadcasts 1997 English Broadcast News Transcripts (Hub-4) - 97 hours of transcripted news broadcasts 1998 HUB5 English Transcripts - transcripts for 20 Callhome English and 20 Switcboard telephone conversations TDT2 Careful Transcription Text - transcripts of broadcasts from 1998
Quick Transcriptions	TDT Pilot Study Corpus - 16,000 news stories TDT2 Multilanguage Text Version 4.0 - news data in Mandarin and English TDT3 Multilanguage Text Version 2.0 - news data in Mandarin and English TDT4 Multilanguage Text Version 1.1: contact LDC for a pre-release copy (LDC catalog no.: LDC2003E21) EARS-only training data releases (contact LDC) LDC2003E12A Fisher Training Speech Part 1 (250 two sided telephone conversations); LDC2003E13A Fisher Quick Transcription Part 1 (250 transcribed two sided telephone conversations); LDC2003E12B Fisher Training Speech Part 2 (250 two sided telephone conversations); LDC2003E13B Fisher Quick Transcription Part 2 (250 transcribed two sided telephone conversations)
Metadata Annotations	EARS-only training data releases (contact LDC) EARS MDE RT-03F Training Corpus (LDC2003E19) EARS MDE RT-03 DevTest and Evaluation Corpus (LDC2003E27)
Conversational Speech	SWITCHBOARD-1 Release 2 - 2,400 two sided telephone conversations Switchboard-2 Phase I - 3,638 five minute conversations Switchboard-2 Phase II - 4,472 five minute conversations Switchboard-2 Phase III - 2,728 two sided telephone conversations CALLHOME American English Speech - 120 telephone conversations up to thirty minutes each CALLFRIEND American English-Non-Southern - 60 telephone conversations five to thirty minutes each CALLFRIEND American English-Southern - 60 telephone conversations five to thirty minutes each Switchboard Cellular Part 1 Audio - 1309 GSM cellular phone calls
Careful Transcriptions	CALLHOME American English Transcripts - 18.3 hours of transcribed speech 1998 HUB5 English Evaluation - 40 sphere files 2001 HUB5 English Evaluation - 60 sphere files Switchboard Cellular Part 1 Transcribed Audio - 250 transcribed files Switchboard Cellular Part 1 Transcription - 250 transcribed files
Pronouncing Lexicon	CALLHOME American English Lexicon - 90,988 lexical entries American English Spoken Lexicon - 50,602 lexical entries of the most common English words
Language Modeling Text	North American News Text Corpus - texts formatted with TIPSTER North American News Text Supplement - texts formatted with TIPSTER English Gigaword ~ 12 GB (over a billion words) of normalized English newswire text EARS-only training data releases (contact LDC) LDC2003E03 TDT4 Multilanguage Text LDC2003E02A TDT4 Multilanguage Supplemental

	Chinese
Broadcast Speech	1997 HUB-4 Broadcast News Evaluation Non English Test - 1 hour each of Spanish and Mandarin news broadcasts 1997 Mandarin Broadcast News Speech (Hub-4NE) - 30 hours of Mandarin news broadcasts TDT2 Mandarin Audio - recorded Mandarin news broadcasts TDT3 Mandarin Audio - recorded Mandarin news broadcasts EARS-only training data releases (contact LDC) LDC2003E02 TDT4 Multilanguage Speech LDC2003E02A TDT4 Multilanguage Supplemental
Careful Transcriptions	1997 Mandarin Broadcast News Transcripts - transcripts for 30 hours of news broadcasts 1997 HUB-4 Broadcast News Evaluation Non English Test - transcripts for 1 hour each of Spanish and Mandarin broadcasts 2001 HUB5 Mandarin Transcripts - 20 Callhome Mandarin telephone conversations
Quick Transcriptions	TDT2 Multilanguage Text Version 4.0 - transcripts for recorded Mandarin news broadcasts TDT3 Multilanguage Text Version 2.0 - transcripts for recorded Mandarin news broadcasts TDT4 Multilanguage Text Version 1.1: contact LDC for a pre-release copy (LDC catalog no.: LDC2003E21)
Conversational Speech	CALLFRIEND Mandarin Chinese-Mainland - 60 Mandarin telephone conversations 5-30 minutes each CALLFRIEND Mandarin Chinese-Taiwan Dialect - 60 Taiwanese dialect telephone conversations 5-30 minutes each CALLHOME Mandarin Chinese Speech - 120 telephone conversations between native Mandarin speakers Hub-5 Mandarin Telephone Speech Corpus - 42 Mandarin telephone conversations Taiwanese Putonghua Speech and Transcripts - 40 speakers using Taiwanese accented Putonghua
Careful Transcriptions	CALLHOME Mandarin Chinese Transcripts - 120 transcripted Mandarin telephone conversations Taiwanese Putonghua Speech and Transcripts - transcripts for 40 speakers using Taiwanese accented Putonghua Hub-5 Mandarin Transcripts - 42 transcripted Mandarin telephone conversations 2001 HUB5 Mandarin Evaluation - 8 hours of transcribed Mandarin conversations
Pronouncing Lexicon	CALLHOME Mandarin Chinese Lexicon - 44,405 words with phonological, morphological and frequency information
Language Modeling Text	Mandarin Chinese News Text - 250 million GB-encoded text characters Chinese Gigaword ~ archive of Chinese newswire text data, totalling over 2.5 million documents or over 1.1 billion words EARS-only training data releases (contact LDC) LDC2003E03 TDT4 Multilanguage Text LDC2003E02A TDT4 Multilanguage Supplemental

	Arabic
Arabic Dialects Transcription Guidelines	Levantine Arabic Transcription Guidelines - (under development)
Broadcast Speech	EARS-only training data releases (contact LDC) LDC2003E02 TDT4 Multilanguage Speech LDC2003E02A TDT4 Multilanguage Supplemental
Careful Transcriptions	Callhome Egyptian Arabic Transcripts Supplement - transcripts for CallHome Egyptian Arabic Speech Supplement
Quick Transcriptions	TDT 3 Arabic Text : contact LDC for a pre-release copy (LDC catalog no.: LDC2002E32 ) TDT4 Multilanguage Text Version 1.1: contact LDC for a pre-release copy (LDC catalog no.: LDC2003E21)
Conversational Speech	CALLHOME Egyptian Arabic Speech - 120 Egyptian Colloquial Arabic telephone conversations CALLFRIEND Egyptian Arabic - 60 telephone conversations between native speaker of Egyptian dialect of Arabic Callhome Egyptian Arabic Speech Supplement - 20 telephone conversations Levantine Arabic QT Training Data Set 1 Speech - 135 phone conversation (up to 10 minutes) of Levantine Arabic speakers (Speech) Levantine Arabic QT Training Data Set 1 Transcripts - 305 phone conversation (up to 10 minutes) of Levantine Arabic speakers (Transcripts) Levantine Arabic QT Training Data Set 2 Speech - 305 phone conversation (up to 10 minutes) of Levantine Arabic speakers (Speech) Levantine Arabic QT Training Data Set 2 Transcripts - 135 phone conversation (up to 10 minutes) of Levantine Arabic speakers (Transcripts)
Careful Transcriptions	CALLHOME Egyptian Arabic Transcripts - transcripts for 120 Egyptian Colloquial Arabic telephone conversations 1997 HUB5 Arabic Evaluation - 20 transcribed conversations 1997 HUB5 Arabic Transcripts - 20 transcribed conversations
Pronouncing Lexicon	Egyptian Colloquial Arabic Lexicon - electronic pronunciation dictionary of Egyptian Colloquial Arabic
Language Modeling Text	Arabic Newswire Part 1 - 2,337 Arabic text data files tagged using TIPSTER Arabic Gigaword ~ archive of Arabic newswire text data, nearly 400 Million words EARS-only training data releases (contact LDC) LDC2003E03 TDT4 Multilanguage Text LDC2003E02A TDT4 Multilanguage Supplemental

tmalec@ldc.upenn.edu

Last modified: Wed Jun 26 09:44:41 2002