Second DIHARD Challenge Development - Eleven Sources LDC2021S10 February 2, 2021 Linguistic Data Consortium 1. Overview =========== This release contains development audio data and annotation for the Second DIHARD Challenge with two exceptions: - the child language domain audio and annotation are distributed in a separate LDC release (LDC2021S11) - due to licensing restrictions, LDC is unable to release the CHiME-5 audio files used for the DIHARD II multichannel condition (track 3 and track 4). To obtain these files, please apply for and obtain the necessary license from University of Sheffield: https://licensing.sheffield.ac.uk/product/chime5 For additional details regarding the challenge, please see the evaluation plan: docs/second_dihard_eval_plan_v1.2.pdf or consult the website: https://dihardchallenge.github.io/dihard2 For instructions on how to reconstruct the complete DIHARD II development set (released during the challenge as LDC2019E31) from this release and LDC2021S11, please see Section 7 of this README. 2. Directory structure ======================= - data/single_channel/flac/ -- single channel condition FLAC files - data/single_channel/rttm/ -- single channel condition RTTM files containing reference diarization - data/single_channel/sad/ -- single channel condition HTK label files containing reference speech segmentation - data/single_channel/uem/ -- single channel condition UEM files for use with "dscore"; one UEM file per domain as well as a file "all.uem" that contains **ALL** scoring regions for **ALL** recordings - data/multichannel/rttm/ -- multichannel condition RTTM files containing reference diarization - data/multichannel/sad/ -- multichannel condition HTK label files containing reference speech segmentation - data/multichannel/uem/ -- multichannel condition UEM files for use with "dscore"; one UEM file per domain as well as a file "all.uem" that contains **ALL** scoring regions for **ALL** recordings - docs/file.tbl -- listing of md5 checksums, sizes, dates, and file names - docs/README.txt -- this file - docs/sources.tbl -- listing of domain, source, and language for each file - docs/second_dihard_eval_plan_v1.2.pdf -- evaluation plan - tools/combine_dev_releases.py -- Python script for reconstructing LDC2019E31 3. File formats =============== 3.1 Audio All single channel condition audio is provided in the form of 16 kHz, 16-bit mono-channel FLAC files. When multiple channels were present in the source audio, these were remixed to a single channel in the delivery version. Similarly, if the source audio was recorded at a sample rate other than 16 kHz, the audio was resampled to 16 kHz. All audio conversion was performed using SoX. Note that the multichannel condition audio is **NOT** provided as part of this release. For these recordings, please apply for and obtain the necessary license from University of Sheffield: https://licensing.sheffield.ac.uk/product/chime5 3.2 Diarization The diarization for each recording is stored as a NIST Rich Transcription Time Marked (RTTM) file. RTTM files are space-delimited text files containing one turn per line, each line containing ten fields: - type -- segment type; should always by "SPEAKER" - file id -- unique identifier for recording that turn is on (e.g., "DH_0001") - channel id -- channel (1-indexed) that turn is on; should always be "1" - turn onset -- onset of turn in seconds from beginning of recording - turn duration -- duration of turn in seconds - orthography field -- should always by "" - speaker type -- should always be "" - speaker name -- name of speaker of turn; should be unique within scope of each file - confidence score -- system confidence (probability) that information is correct; should always be "" - signal lookahead time -- should always be "" While the RTTM format allows for a single file to contain turns from multiple recordings, the DIHARD RTTM files always contain turns from a single recording; e.g., "DH_0001.rttm" only contains turns from recording "DH_0001". 3.3 Speech segmentation For each recording a reference speech segmentation is generated by merging all overlapping speaker turns. These segmentations are stored as HTK label files ending in the ".lab" extension. Each file contains one speech segment per line, each line containing three tab-delimited fields: - onset -- onset of speech segment in seconds from begining of recording - offset -- offset of speech segment in seconds from beginning of recording - label -- label of segment; always "speech" 3.4 Scoring regions The scoring regions for each recording are specified by un-partitioned evaluation map (UEM) files. These files contain one line per scoring region, each line consisting of four space-delimited fields: - file id -- unique identifier for recording that scoring region is on; (e.g., "DH_0001") - channel -- channel (1-indexed) that scoring region is on - onset -- onset of scoring region in seconds from beginning of recording - offset -- offset of scoring region in seconds from beginning of recording For each audio condition (single channel and multichannel) there is a single UEM listing **ALL** scoring regions for **ALL** recordings. For the single channel condition, this is: data/single_channel/uem/all.uem while for the multichannel condition this is: data/multichannel/uem/all.uem Additionally, UEMs have been provided for each individual domain of the single channel condition: data/single_channel/uem/audiobooks.uem data/single_channel/uem/broadcast_interview.uem data/single_channel/uem/clinical.uem ... 4. Recording metadata ====================== 4.1 sources.tbl Metadata for each recording is stored in "docs/sources.tbl". This file is a tab-delimited table containing one recording per line, each line having the following 4 fields: - file id -- unique identifier for recording - lang -- predominant language of recording; ISO 639-3 language code - domain -- domain of recording; see section 4.2 for details - source -- source of recording; see section 4.3 for details 4.2 Recording domains Recordings are drawn from the following 12 domains: - AUDIOBOOKS -- amateur audiobooks - BROADCAST-INTERVIEW -- radio interviews - CHILD -- child language acquisition recordings* - CLINICAL -- Autism Diagnostic Observation Schedule (ADOS) interviews - COURT -- courtroom recordings - DINNER -- dinner party - MAPTASK -- map tasks - MEETING -- meeting speech - RESTAURANT -- conversational speech recorded during lunches in restaurants - SOCIO-FIELD -- sociolinguistic interviews recorded in the field - SOCIO-LAB -- sociolinguistic interviews recorded in laboratory setting - WEBVIDEO -- amateur video from sites such as YouTube * All data from CHILD domain is distributed as LDC2021S11. 4.3 Recording sources Recordings are drawn from the following 12 sources: - ADOS -- unpublished ADOS interviews recorded at CHOP - CHIME -- CHiME-5 recordings - CIR -- unpublished LDC collection of restaurant conversation - DCIEM -- DCIEM map task (LDC96S38) - LIBRIVOX -- audiobook recordings from LibriVox - MIXER6 -- interviews from MIXER6 (LDC2013S03) - RT04S -- meeting speech from 2004 Spring NIST Rich Transcription (RT-04S) dev (LDC2007S11) and eval (LDC2007S12) releases - SCOTUS -- 2001 U.S. Supreme Court oral arguments from OYEZ project - SEEDLINGS -- child language recordings collected as part of SEEDLingS* - SLX -- sociolinguistic interviews drawn from SLX (LDC2003T15) - VAST -- web video collected as part of the Video Annotation for Speech Technologies (VAST) project - YOUTHPOINT -- unpublished corpus of student-led radio interviews conducted for YouthPoint, a 1970s radio program * All SEEDLingS data is distributed as LDC2021S11. 4.4 Additional information For full information regarding each source and domain, please consult Appendix A from the evaluation plan. 5. File table ============= Expected sizes, modification times, and MD5 checksums for all files within the "data/" directory are recorded in "docs/file.tbl". This is a tab-delimited table containing one file per line, each line having the following 4 fields: - checksum -- MD5 checksum of file - size -- size of file in bytes - datetime -- last modification date in YYYY-MM-DD_HH:MM:SS format - path -- path to file relative to root of release directory 6. Relationship to DIHARD II tracks =================================== Tracks 1 and 2 of DIHARD use the single channel condition audio and annotations; that is, the contents of: data/single_channel Tracks 3 and 4 use the multi-channel condition annotations: data/multichannel As a reminder, LDC is **NOT** distributing the audio for the multichannel data. To obtain the audio, please contact the CHiME organizers: https://licensing.sheffield.ac.uk/product/chime5 7. Reconstructing LDC2019E31 ============================ To reconstruct the full DIHARD II development set (LDC2019E31) from this release and LDC2021S11, please run the script "tools/combine_dev_releases.py" as follows: python tools/combine_dev_releases.py 11-sources-dir seedlings-dir combined-dir where: - 11-sources-dir -- path to LDC2021S10 (this release) - seedlings-dir -- path to LDC2021S11 - combined-dir -- output path for reconstructed development set Please note that this script does require that "pandas" is installed in your Python environment. 8. Contacts =========== If you have questions about this data release, please contact the following LDC personnel: Neville Ryant