README FILE FOR: LDC2026S09 CORPUS TITLE: 2012 NIST Speaker Recognition Evaluation Test Set AUTHORS: Craig Greenberg, Alvin Martin (NIST) David Graff, Kevin Walker, Stephanie Strassel, Karen Jones (LDC) 1.0 Introduction This release contains the English language test segments and trial and answer keys for the 2012 NIST Speaker Recognition Evaluation (SRE12) test set and additional segments intended for modeling. There are 177,792 audio files in the package, comprising 10,321 hours of audio data. The Speaker Recognition Evaluation series is a recurring set of evaluations conducted by NIST that are intended to be of interest to researchers working on the general problem of text independent speaker recognition. For SRE12, the speakers had been represented in prior SRE collections. The basic task of the SRE evaluation is to determine whether a specified speaker is talking in a given segment of speech. 2.0 Data Sources The modeling and test data in this package originate from several Mixer style collections involving recorded telephone conversations and or in-person recorded studio sessions. For the telephone calls, recruited speakers dialed into a dedicated telephone number and were connected to another participant via a robot operator to talk for 10-12 minutes on a topic that was announced at the beginning of the recorded call. The SRE12 evaluation test set originates from two corpora collected by the LDC: a) Mixer 7, a dataset collected for IARPA's Biometric Exploitation Science and Technology (BEST) program consisting of multi-microphone recorded speech and conversational telephone speech (CTS). The multi-microphone portion involved recording speakers with 14 distinct microphones installed identically in two multi-channel audio recording rooms at LDC. Each session was guided by collection staff using prompting and recording software to conduct the following activities: (1) repeat questions (less than one minute); (2) informal conversation (typically 15 minutes); (3) transcript reading (15 minutes); and (4) up to three telephone calls under varying conditions (10 minutes). For the CTS portion of Mixer 7, the collection protocol was in line with prior Mixers as described above. b) Remix, a collection consisting solely of conversational telephone speech collected under the same protocol as past Mixer collections, as described above. The evaluation test segments drawn from both of these collections were restricted to English. Some target speakers for SRE12 were present in the test data for two prior SRE evaluations, namely SRE08 and SRE10, while others had not appeared in any prior SRE evaluations. The modeling data originates from several Mixer collections. Below is a inventory of all source collections used in this package (both test set and modeling data): DATASET COLLECTION YEARS LDC CATALOG ID USED IN NIST SRE/LRE EVAL Mixer 3 2005-2007 LDC2023S02 SRE2006, LRE2007 Mixer 4/5 2007 LDC2020S03 SRE2008 Mixer 6 2009-2010 LDC2013S03 SRE2010 Mixer 7 English Speech 2011-2012 LDC2025S08 SRE2012 Mixer 7 Spanish Speech 2011-2012 LDC2023S04 SRE2012 Remix 2011-2012 LDC2023S09 SRE2012 The datasets listed above are available via the LDC Catalog: https://catalog.ldc.upenn.edu/. The above datasets were collected by LDC, and, in the case of Mixers 4 and 5 in partnership with the International Computer Science Institute (ICSI) at the University of California, Berkeley. 3.0 Languages All test segments are in English and the segments intended for modeling are mostly in English, but some other languages are present, as follows, since some Mixer participants recorded calls in both English and another language: Mixer 3 - English, Korean, Cantonese, Russian, Tagalog, and Mandarin Mixer 4 and 5 - English Mixer 6 - English Mixer 7 - Spanish and English Remix - English 4.0 Directory Structure /data /model_segs /sp06 /sp08 /10sec /long /short2 /short3 /summed /sp10 /10sec_0 /10sec_1 /interview_3min /interview_8min /phonecall_mic /phonecall_tel /summed /test_segs /mic_int /iaaakw-idrzps /idrzri-ihkomx /ihkonl-ilewbs /ilewdr-ioxzpy /ioyaky-ispouv /ispoya-iwhxry /iwhyoh-izzzqg /mic_phn /maaawa-mffrat /mffrgp-mkhlaz /mkhlsm-mpmjge /mpmjkj-muubmh /muudkz-mzzytq /sum_phn /saaant-sinzgb /sinzja-srfykb /srfylq-szzzzx /tel_phn /taaaho-tfdwxt /tfdwyy-tkjxod /tkjxwc-tpovkh /tpowig-tuvbrd /tuvbrj-tzzzrk /docs file.tbl NIST_SRE12_evalplan-v17-r1.pdf NIST_SRE12_segment_table.v2.csv /NIST_SRE12_target_speaker_tables_v2.2 NIST_SRE12_target_speaker_files_map.v2.2.txt NIST_SRE12_target_speaker_single_file_per_ldcid_map.v2.2.txt NIST_SRE12_target_speaker_speakertable.v2.csv /trialkeys NIST_SRE12_core_trial_key.csv NIST_SRE12_extended_trial_key.csv NIST_SRE12_summed_trial_key.csv NIST_SRE12_supplemental_trial_key.csv README.txt -- this file 5.0 Data This release contains: 130,844 test segments (6623.01 hours of audio) 46,948 model segments (3704.65 hours of audio) The audio is presented in uncompressed NIST SPHERE format. These are organized into four categories (where "PTN" refers to "Public Telephone Network"): mic_int -- in-room microphone recordings of interview conversation mic_phn -- in-room microphone recordings of telephone conversation sum_phn -- PTN recordings with 2 call-sides combined into 1 channel tel_phn -- PTN recordings with 2 call-sides presented as separate channels The two "mic_*" sets contain 2-channel, 16-bit PCM, 16-KHz audio, and the "sum_phn" and "tel_phn" sets contain 8-bit Mu-law, 8-KHz audio. Test segments of three different durations were cut from each recording. The quantities for each duration are given below: Secs #segs 30 32788 100 32788 300 65268 An additional 46,948 files intended for modeling are included in this package. 6.0 Metadata 6.1 docs/file.tbl This file provides checksum, file size, date/time and path information for each audio file. 6.2 docs/NIST_SRE12_evalplan-v17-r1.pdf This file is a copy of the 2012 NIST Speaker Recognition Evaluation Plan. 6.3 docs/NIST_SRE12_segment_table.v2.csv This file provides the following segment metadata: segmentid - NIST segment ID ldcid - LDC segment ID speech_style - variables include: interview, phonecall channel_type - variables include: mic_CH02, mic_CH04, mic_CH07, mic_CH12, summed, tel duration_in_secs - variables include: 30,100,300 alteration_type - Added noise, variables include: 15db_noise_babble, 15db_noise_hvac, 6db_noise_babble, 6db_noise_hvac, 6db_noise_singlespeaker, no_alteration error_status - variables include: error, no_error An example is given below: segmentid iaaakw_sre12 ldcid 20070420_133152_PHL_103187 speech_style interview channel_type mic_CH02 duration_in_secs 100 alteration_type no_alteration error_status no_error 6.4 /docs/NIST_SRE12_target_speaker_tables_v2.2 This directory provides speaker information. 6.4.1 NIST_SRE12_target_speaker_2_files_map.v2.2.txt This file provides a mapping of speaker-id to relative paths for all audio segments in which the speaker is present. This may include repeated speech over different channels or for different durations. 6.4.2 NIST_SRE12_target_speaker_2_single_file_per_ldcid_map.v2.2.txt This file provides a mapping of speaker-id to relative paths for all audio segments in which the speaker is present such that there is no repeated speech. When selecting which segments to include in the map, first preference was for longer duration, then phone call speech style over interview speech style, then telephone channel over microphone channel, then for close talking microphones over distant microphones. Least preference was given to summed channel phone calls. Ties were broken by lexicographic ordering of filenames. 6.4.3 NIST_SRE12_target_speaker_speakertable.v2.csv This file provides demographic information (where available) about each speaker; the type of information provided includes: speakerid sex year_of_birth years_of_education age_learned_english_as_a_second_language native_language other_languages_spoke occupation country_born state_born city_born country_raised state_raised city_raised ethnicity is_a_smoker height_cm weight_kg 6.5 /docs/trialkeys This directory contains information about segments in the required "core" evaluation condition and three optional "extended", "summed", and "supplemental" conditions. The core test condition involved two-channel test segments in a "small" trial (limited to 2 million trials). The extended and supplemental condition involved two-channel test segments in "large" trials (limited to 100 million trials), while the summed condition involved segments formed from summing the two conversational channels to be used in a small trial. Each of the following files in this directory NIST_SRE12_core_trial_key.csv NIST_SRE12_extended_trial_key.csv NIST_SRE12_summed_trial_key.csv NIST_SRE12_supplemental_trial_key.csv contains the following fields: speakerid relative_path_to_test_segment side - side of call, values: a, b targettype - informs whether speaker was a target or not, variables: target, nontarget is_scored - informs whether segment was scored, variables: is_scored, not_scored An example is given below: speakerid 100304 relative_path_to_test_segment data/tel_phn/taaaho-tfdwxt/tacjol_sre12.sph side b targettype nontarget is_scored scored Further information about core, extended, summed and supplemental training/test conditions can be found in the 2012 NIST Speaker Recognition Evaluation Plan included in this package as /NIST_SRE12_evalplan-v17-r1.pdf 7.0 Known Issues The following two issues were identified at the time of the SRE12 evaluation: some sets of evaluation files with distinct paths have identical content; and some 2-channel files have speech on only one channel while some 1-channel files contain no speech. The answer-key data handles such issues whenever data of this kind is referenced in the list of trials. There is also a single invalid modeling audio file: /model_segs/sp08/short3/ffttr.sph 8.0 Copyright Information © 2012 Trustees of the University of Pennsylvania 9.0 CONTACTS Craig Greenberg, NIST README created by Karen Jones on May 17, 2016 updated by David Graff and Karen Jones on May 30, 2023 updated by Karen Jones on April 18,2025