Cerru02/LombardGrid-Retrieval

Lombard GRID Cross-Modal Utterance Retrieval

Four MTEB/MOEB utterance-retrieval tasks derived from the Lombard GRID corpus: audio-to-video, video-to-audio, frontal-to-profile video, and plain-to-Lombard video+audio. Relevance always requires the same utterance recording or the same speaker-and-sentence pair; speaker identity alone is never relevant.

The source paper introduced the corpus but did not define these retrieval tasks. Relevance is derived from native recording, speaker, sentence-code, condition, and camera-view labels.

Frozen evaluation protocol

• Speakers: 54 (s2 through s55; s1 was excluded by the source)
• Selected sentence codes: 540 (10 per speaker)
• Published media recordings: 1080
• Matching-task recordings: 540 (5 plain and 5 Lombard per speaker)
• Selection seed: mteb-lombard-grid-retrieval-v2

For every speaker, eligible sentence codes present in both conditions and all three media streams are ordered by SHA-256 over the fixed seed, speaker, and sentence code. The first ten valid pairs are retained. Five plain and five Lombard recordings per speaker are then selected deterministically for the three same-recording tasks. Opaque IDs are published; speaker, condition, sentence code, transcription, and other text are not encodable dataset features.

Task	Queries	Corpus	Qrels	Positives/query
Audio to frontal/profile video	540	1,080	1,080	2
Frontal/profile video to audio	1,080	540	1,080	1
Frontal video to profile video	540	540	540	1
Plain frontal video+audio to Lombard frontal video+audio	540	540	540	1

Source audit and release discrepancies

The official JSON archive contains 5,340 rows, while each media archive contains 5,390 files. All three media archives have exactly the same recording filenames. The media restore 50 s31 recordings absent from JSON. s51 has only 40 plain and 50 Lombard files, rather than the advertised 50 of each. s14 has 51 plain and 49 Lombard rows, including a duplicate prwr3a recording key. The JSON metadata labels 69 recordings as WRONG. Media filenames mark 70: the additional one is among the 50 media-only s31 recordings. All 70, plus the one legacy-named duplicate, are excluded from selection. Four frontal MOV files lack a moov atom and are corrupt: s32_l_pwip9p, s32_p_bwwj2n, s33_l_pwajza, and s33_p_sgwq2s.

The deterministic selection required 0 fallback(s). The selected protocol remains exactly balanced after exclusions.

Media processing and validation

The released MOV files are already H.264/yuv420p visual-only clips, so 2,160 selected videos are copied without transcoding. 0 video(s) required a lossless video-stream remux to remove embedded audio. Separate source WAV files are copied without transcoding.

All 1080 frontal videos, 1080 profile videos, and 1080 audio files were decoded completely with TorchCodec during construction. Matching source filename stems establish audiovisual correspondence. Selected video/audio duration differences were checked and paired decoding succeeded. The source paper reports correlation-based audiovisual alignment before utterance extraction; this construction did not independently verify signal-level synchronization.

The shared configs avoid storing the same media repeatedly across four tasks:

• media/test: opaque recording and modality IDs, frontal video, profile video, audio
• matching_utterances/test: opaque audio, frontal-video, and profile-video IDs
• condition_pairs/test: opaque plain- and Lombard-recording IDs

The MTEB task loaders project these configs into the standard query, corpus, and qrel structures without exposing source labels to embedding models.

Evaluation

The primary metric for all four tasks is nDCG@10. Other standard MTEB retrieval metrics are reported as secondary results. These protocols are newly derived, so there is no source-paper retrieval score to reproduce.

License and attribution

The source corpus is released under the Creative Commons Attribution 4.0 International license. Please attribute the original authors and cite the paper below. The MTEB task metadata and this card both use cc-by-4.0.

Responsible use

The source contains identifiable faces and voices. Although these tasks evaluate utterance correspondence rather than biometric identification, the representations may still encode identity and demographic attributes. Users should consider participant consent, privacy, applicable biometric-data law, subgroup behavior, and downstream misuse before training or deploying related systems.

Citation

@article{alghamdi2018corpus,
  author = {Alghamdi, Najwa and Maddock, Steve and Marxer, Ricard and Barker, Jon and Brown, Guy J.},
  title = {A corpus of audio-visual Lombard speech with frontal and profile views},
  journal = {The Journal of the Acoustical Society of America},
  volume = {143},
  number = {6},
  pages = {EL523--EL529},
  year = {2018},
  doi = {10.1121/1.5042758},
}