Fall 2026 LDC data scholarship program
New publications:
2012 NIST Speaker Recognition Evaluation Test Set
CALLHOME American English Second Edition
CALLHOME American English Lexicon (PRONLEX) Second Edition
___________________________________________________________________
New publications:
2012 NIST Speaker Recognition Evaluation Test Set was developed by LDC and NIST and contains 10,321 hours of English conversational telephone speech and in-person recorded studio sessions for evaluation and modeling, along with answer keys, trial files and documentation from the NIST-sponsored 2012 Speaker Recognition Evaluation (SRE12). SRE12 introduced a revised evaluation structure in which training data for target speakers was drawn from prior SRE corpora developed by LDC and was provided in advance of the evaluation period.
Test data was drawn from Mixer 7 English Speech (LDC2025S08) and REMIX Telephone Collection (LDC2023S09). Those datasets also provided segments for modeling data; other modeling segments were drawn from Mixer 3 Speech (LDC2023S02), Mixer 4 and 5 Speech (LDC2020S03), and Mixer 6 Speech (LDC2013S03). The test data contains English speech only; some non-English speech is contained in modeling segments.
This release is comprised of 130,844 test segments, specifically, 83,778 call segments and 47,066 interview segments. Modeling data consists of 46,948 segments.
2026 members can access this corpus through their LDC accounts. Non-members may license this data for a fee.
CALLHOME American English Second Edition was developed by LDC and contains 56 hours of speech from 120 unscripted telephone conversations between native American English speakers. This publication is a re-release of the original CALLHOME American English collection, combining CALLHOME American English Speech (LDC97S42) and CALLHOME American English Transcripts (LDC97T14), with additional transcription and updated directory structure, file formats, and documentation.
This release contains the 120 telephone conversations published in CALLHOME American English Speech which represented training and development data and a subset of evaluation data. Participants spoke on topics of their choice in a single telephone call lasting up to 30 minutes. Calls were manually audited for gender, language, recording quality, channel characteristics, dialect, and region. For this second edition, all audio was converted from SPHERE files to FLAC format, and the original training/development/evaluation partitioning was removed.
This release also features revised transcripts conforming to updated LDC transcription guidelines that addressed normalization of annotation formats, standardization of speaker-produced and background noises, application of foreign-language marking, whitespace cleanup, and corrections and consistency fixes.
The CALLHOME series consists of telephone conversations and transcripts developed by LDC and Rutgers, The State University of New Jersey, in support of research in speaker identification, language identification and related technologies. Languages in the series include American English, Egyptian Arabic, German, Japanese, Mandarin Chinese, and Spanish.
2026 members can access this corpus through their LDC accounts. Non-members may license this data for a fee.
CALLHOME American English Lexicon (PRONLEX) Second Edition was developed by LDC and contains 90,988 English words with citation-form pronunciations. This second edition updates file formats, directory structure and documentation. The first edition is available as CALLHOME American English Lexicon (PRONLEX) (LDC97L20).
The words in the lexicon were derived from Wall Street Journal text used in the continuous speech recognition publication series CSR-1 WSJ0 Complete (LDC93S6A), transcripts from the Switchboard telephone collection (LDC97S62), and transcripts representing unscripted telephone conversations between native American English speakers contained in CALLHOME American English Second Edition (LDC2026S08).
PRONLEX transcription is a phonemic transcription system designed to support speech recognition by providing a consistent and simplified representation of how words are pronounced in standard American English that allows variation to be generated later to avoid listing many pronunciation variations for each word. This single systematic base form can be expanded through rules or modeling. The transcription was created using a modified ARPABET phoneme set.
The lexicon contains three tab-separated information fields: (1) word: orthographic representation of word; (2) pron: transcribed citation-form pronunciations using modified ARPABET phoneme set; and (3) comments: (OPTIONAL) comment on the entry. It is presented as a tab-delimited TSV file encoded in UTF-8 format and includes a pronunciation dictionary derived from the lexicon in UTF-8 encoded CMUdict format.
2026 members can access this corpus through their LDC accounts provided they have submitted a completed copy of the special license agreement. Non-members may license this data for a fee.
