JeffeLI/sre24_train

SRE24 Training Data (WeSpeaker Shard Format)

This dataset provides the fixed training set used in the NIST 2024 Speaker Recognition Evaluation (SRE24) together with the SRE24 development (DEV) data, packaged and reorganized into the WeSpeaker shard (tar) format for direct use with the WeSpeaker speaker recognition toolkit.

Paper Reference: C. Greenberg et al., "The 2024 NIST Speaker Recognition Evaluation", Proc. Interspeech 2025, Rotterdam, The Netherlands, Aug 2025. [PDF]

---

Dataset Structure

Split	Size	Files	Description
sre	~300 GB	646 shard tar files	SRE24 Fixed Training Set (SRE CTS Superset, SRE16, SRE21, SRE24 DEV, JANUS, etc.)
sre24_dev	~1.5 GB	16 files (incl. shards)	SRE24 Official Development Set (DEV)
sre24_eval	~21.7 GB	48 files (incl. shards)	SRE24 Evaluation Test Set (EVAL)

sre24_train/
├── sre/                          # Fixed training set (primary training data)
│   ├── shards/                   # ~644 tar shards, each containing utterances
│   ├── shard.list                # List of shard tar paths (for WeSpeaker training)
│   └── utt2spk                   # Utterance-to-speaker mapping
│
├── sre24_dev/                    # Development / validation set
│   ├── shards/
│   ├── shard.list
│   ├── raw.list
│   ├── utt2spk
│   ├── spk2utt
│   ├── wav.scp                   # Original Kaldi-format wav scp
│   ├── check.sh
│   └── trail_*                   # Trial lists:
│         trail_key               #   key / standard trial
│         trail_total
│         trail_afv_afv           #   cross-condition trials
│         trail_afv_cts
│         trail_cts_afv
│         trail_cts_cts
│
└── sre24_eval/                   # Evaluation (EVAL) test set
    ├── shards/
    ├── shard.list
    ├── utt2spk
    ├── wav.scp
    └── trail_key

---

Background (from SRE24 Paper, Section 3)

3.1 Training Set (Fixed Condition)

The fixed training condition was mandatory for all SRE24 participants and provides a common, reproducible benchmark that decouples training-data effects from model/algorithm effects.

The fixed training set in sre/ is composed of the following corpora (made available by the LDC):

Corpus	LDC Catalog	Role
SRE CTS Superset	LDC2021E08	Primary CTS (Conversational Telephone Speech) training
SRE16 Evaluation Test Set	LDC2019S20	Addl. CTS / cross-lingual data
SRE21 Evaluation Test & Dev Set	LDC2024E10	CTS + Audio-from-Video (AfV) data
SRE24 Evaluation Dev Set: Data	LDC2024E12	SRE24 DEV audio/visual data
SRE24 Evaluation Dev Set: Annotations	LDC2024E34	SRE24 DEV annotations/trials
JANUS Multimedia Dataset	LDC2019E55	Multimodal (audio + video) data

In the open training condition (optional), participants could add external data on top.

3.2 Development & Test Sets (DEV / EVAL)

Both DEV (sre24_dev/) and TEST/EVAL (sre24_eval/) are drawn from the TELVID corpus — a multilingual, multimodal corpus collected outside North America (English, French, Tunisian Arabic) by the LDC. TELVID contains:

• CTS (Conversational Telephone Speech): A-law 8 kHz SPHERE files.
• AfV (Audio from Video): 16-bit 16 kHz FLAC files.
• Video recordings + still "selfie" images (not included in this audio-only release).

Key data statistics (from the SRE24 paper Table 1):

Track	Subset	Speakers (M/F)	Enroll (1/3-seg)	Test segs	Target trials	Non-target trials
Audio	DEV	10/10	1023/116	2077	110,738	1,064,760
Audio	TEST (eval)	124/163	14,408/1590	29,487	1,565,121	6,930,082
Visual	DEV	10/10	20/–	455	455	4,095
Visual	TEST	124/163	287/–	6,848	6,848	783,560
Audio-Visual	DEV	10/10	1023/116	455	25,356	233,216
Audio-Visual	TEST	124/163	14,408/1590	6,848	382,274	1,607,411

SRE24 introduced three novel features:

1. Variable-duration enrollment segments: ~10s, ~30s, ~60s (vs. fixed 60s previously).
2. Shorter test segments: 5s–60s range (vs. previous 10s–60s).
3. Multi-person segments: Enrollment/test segments may contain more than one speaker; diarization marks are supplied for target speakers in such enroll segments.

---

WeSpeaker Shard Format

This dataset uses the WeSpeaker shard (tar) layout produced by make_shard_list.py. This format groups utterances into fixed-size .tar shards for fast sequential / distributed training.

Shard Tar Contents

Each shards/shards_XXXXXXXXXX.tar contains one .wav (or .flac / .sph) file and one .spk (speaker label) file per utterance:

shards_000000000.tar/
    ├── utt_id_000001.wav      # Raw audio bytes (wav/flac/sphere)
    ├── utt_id_000001.spk      # Speaker label stored as UTF-8 text
    ├── utt_id_000002.wav
    ├── utt_id_000002.spk
    └── ...

What make_shard_list.py Does (Reference Workflow)

Reference: https://github.com/wenet-e2e/wespeaker/blob/master/tools/make_shard_list.py

python wespeaker/tools/make_shard_list.py \
    --num_utts_per_shard 1000 \     # ~1k utts per tar shard
    --num_threads 8 \               # parallel packing
    --prefix shards \               # shard filename prefix
    --shuffle \                     # shuffle before sharding
    wav.scp utt2spk \               # inputs: kaldi-format wav list + utt2spk
    exp/shards_dir/ exp/shards.list # outputs: shard folder + shard list

Essential logic of make_shard_list.py:

1. Parses Kaldi-style wav.scp (utt_id → wav_path or pipe) and utt2spk (utt_id → spk_id).
2. Groups (key, spk, wav) into chunks of --num_utts_per_shard.
3. Writes each chunk into {prefix}_{:09d}.tar using Python tarfile, adding two members per utterance:
   • key + ".wav" (or .flac / .wma / ... depending on the original suffix) → raw audio bytes.
   • key + ".spk" → spk encoded as UTF-8 bytes.
4. Supports optional VAD via --vad_file (applies VAD trimming before packing).
5. Emits a shard list file (shard.list), a text file with one absolute/relative tar path per line, used directly by WeSpeaker's data loader.

---

How to Use This Dataset in WeSpeaker

1. Clone / install WeSpeaker

git clone https://github.com/wenet-e2e/wespeaker.git
cd wespeaker
pip install -r requirements.txt

2. Point WeSpeaker's training config at this dataset

The directory layout mirrors exactly what WeSpeaker expects. The shard.list file can be fed directly into WeSpeaker's ShardDataset:

# Example: wespeaker/conf/*.yaml
dataset:
  train:
    type: ShardDataset
    shards_list: /path/to/sre24_train/sre/shard.list   # <-- THIS DATASET
    num_workers: 8
    shuffle: True
    ...
  cv:
    type: ShardDataset
    shards_list: /path/to/sre24_train/sre24_dev/shard.list
    ...

3. Re-sharding / regenerating shards (optional)

If you need to re-shard (different --num_utts_per_shard), start from the original wav.scp + utt2spk:

python wespeaker/tools/make_shard_list.py \
    --num_utts_per_shard 2000 \
    --num_threads 16 \
    --shuffle \
    sre24_train/sre24_eval/wav.scp \
    sre24_train/sre24_eval/utt2spk \
    sre24_train/sre24_eval/shards \
    sre24_train/sre24_eval/shard.list

4. Scoring trials (DEV / EVAL)

Use the trail_* files in sre24_dev/ and sre24_eval/ together with your favourite scorer (e.g. WeSpeaker's wespeaker/bin/extract_embedding.py + wespeaker/bin/score_plda.py or plain cosine scoring).

# Example: cosine scoring with wespeaker
python wespeaker/examples/voxceleb/v2/local/score.py \
    --scoring cosine \
    --trial_file sre24_train/sre24_dev/trail_key \
    --emb_scp exp/xvector.scp \
    --output_file exp/scores

---

Content Warnings & Licensing

⚠️ VERY IMPORTANT — Redistribution Restrictions

This Hugging Face repository contains repackaged speech data whose original source corpora are distributed by the Linguistic Data Consortium (LDC) at the University of Pennsylvania.

Under the LDC User Agreement (both the standard Membership Agreement and the Non-Member User Agreement), users who receive LDC data are expressly prohibited from publishing, retransmitting, disclosing, copying, reproducing or redistributing LDC Databases to anyone outside of their own Research Group:

"Unless explicitly permitted herein, User shall not otherwise publish, retransmit, disclose, display, copy, reproduce or redistribute the LDC Databases to others outside of User's Research Group. User shall have no right to copy, redistribute, transmit, publish or otherwise use the LDC Databases for any other purpose." — LDC User Agreement for Non-Members, §¶ Use & Redistribution clause.

LDC further confirms (see e.g. LDC June 2024 Newsletter):

"LDC data cannot be shared outside the member/licensing organization. LDC reserves the right to deactivate user accounts if any suspicious activity is detected."

What this means for you as a downloader of this repository:

1. Access is conditional on you already having a valid LDC license / membership for ALL of the individual corpora listed below. If you are not already an LDC member / licensee for these corpora, you must obtain the licenses directly from LDC before using any data from this repo.
2. Do NOT further re-share / mirror / re-upload the contents of this repository (or any derivatives) to any public site, cloud bucket, torrent, or LLM training pool. Doing so would violate the upstream LDC license and may expose you to legal liability.
3. Use is limited to non-commercial linguistic education, research and technology development only. Any commercial product, commercial feature, or commercial evaluation derived from these corpora requires the organization to become an LDC For-Profit member prior to product release, per LDC policy.

Source Corpora and their LDC Catalog IDs

This repository draws from the following LDC corpora (see SRE24 paper §3.1). To use this data legally you must have licensed all of them from the LDC:

Corpus / Component	LDC Catalog ID	Link
SRE CTS Superset	LDC2021E08	https://catalog.ldc.upenn.edu/LDC2021E08
SRE16 Evaluation Test Set	LDC2019S20	https://catalog.ldc.upenn.edu/LDC2019S20
SRE21 Evaluation Test and Dev Set	LDC2024E10	https://catalog.ldc.upenn.edu/LDC2024E10
SRE24 Evaluation Dev Set: Data	LDC2024E12	https://catalog.ldc.upenn.edu/LDC2024E12
SRE24 Evaluation Dev Set: Annotations	LDC2024E34	https://catalog.ldc.upenn.edu/LDC2024E34
JANUS Multimedia Dataset	LDC2019E55	https://catalog.ldc.upenn.edu/LDC2019E55
TELVID (SRE24 DEV + TEST audio+video)	(distributed via LDC2024E12/E34 & SRE24 eval package)	https://www.ldc.upenn.edu/

(SRE CTS Superset itself further incorporates LDC97S62, LDC98S75, LDC99S79, LDC2002S06, LDC2001S13, LDC2004S07, LDC2021R03, LDC2020S03, LDC2013S03, LDC2013S05 — i.e. Switchboard, Mixer 3/4-5/6, Greybeard — see Sadjadi 2021, NIST SRE CTS Superset, Table 1.)

How to license the data from LDC

• Academic institutions: LDC Non-Member license (fee per corpus) or Academic Membership (annual fee, unlimited access to most corpora).
• For-profit companies: LDC For-Profit Membership is a prerequisite for any commercial use and to obtain commercial licenses.
• Start page: https://www.ldc.upenn.edu/data-management/using/licensing
• Licensing page PDF (non-member): https://catalog.ldc.upenn.edu/license/ldc-non-members-agreement.pdf

Intended Use & Ethics

• Allowed: Non-commercial speaker-recognition / speaker-verification research (training, validation, evaluation benchmarking, publication).
• Forbidden (without an LDC commercial license): Commercial product development, testing commercial product features, model training that will ship in commercial products.
• Absolutely forbidden: Any re-distribution outside your licensed Research Group (see above); any form of biometric surveillance, mass speaker profiling, or deanonymisation targeting real individuals who are party to LDC consent agreements.
• Data was collected by LDC with informed consent. SRE24 DEV and TEST speakers are explicitly disjoint. Citation format suggestions from LDC should be used in any scholarly publication: cite the individual LDC corpora, the NIST SRE24 overview paper, and (if applicable) the WeSpeaker toolkit.

---

Acknowledgements

• NIST SRE24 organizers: Craig Greenberg, Lukas Diduch, Audrey Tong, Elliot Singer, Trang Nguyen, Robert Dunn, Lisa Mason, Beth Matys (NIST / MIT-LL / DoD).
• LDC for curating and distributing the source corpora including SRE CTS Superset, TELVID, JANUS, SRE16/21/24 DEV.
• WeSpeaker team for the make_shard_list.py shard utility and end-to-end training framework.
• Dataset card compiled from the SRE24 overview paper (Greenberg et al., Interspeech 2025).