sammyliu/audio-jev-rlcd-v1

Audio Jev RLCD v1

Causal audio windows and runtime action questions for reinforcement learning and evaluation. Audio is supplied once; each question receives a categorical label distribution. This corpus contains gold/source annotations. Teacher probabilities and policy-generated labels are not substituted for gold.

Contents

Split	Decisions	Source windows	Stored audio blobs	Natural source windows	Audio hours	Windows over 120 s
train	168,494	71,952	71,952	59,952	525.44	230
dev	14,873	8,900	8,900	8,900	36.46	27
calibration	15,425	8,278	8,278	8,278	46.54	34
test	82,331	33,944	33,854	33,944	341.20	382
external_test	9,998	9,998	9,998	4,999	23.54	0

Source windows count distinct selected source offsets/IDs; stored audio blobs additionally collapse exact decoded waveform duplicates within a split. Hours sum supplied windows, including overlapping windows; they are not independent source-recording hours. The external_test diagnostic is stored under a separate corpus root because its source collections can overlap development/training collections. It is never a selection split.

Labels and observation boundaries

• speech_activity: any human speech intersects the final 200 ms of supplied audio.
• speaker_count: distinct audible source speaker identities anywhere in the supplied window, with labels 0–9 and 10+. Silent attendees cannot be counted.
• turn_complete: the supplied Smart Turn semantic endpoint annotation, complete/incomplete.
• floor_transition: observed same-speaker resumption, different-speaker onset, simultaneous onset, or no onset in the following two seconds. This is a future-outcome proxy, not a gold semantic completion or backchannel judgment.
• next_speaker: the next observed speaker, identified by order of first audible appearance, a new voice, or nobody. Ambiguous input/onset ordering is excluded. Its candidate list contains the source-annotated audible voices, so candidate-set size supplies an oracle speaker inventory; this is not unconstrained speaker discovery. The separate counting task always uses fixed 0–10+ labels.

All model-visible audio is physically cropped before any future-derived target is attached. Future segments are used only for source annotations. Inputs contain no transcripts, future audio, source labels, teacher rationales, or provenance strings. Targets and provenance must remain outside model context. Every categorical target sums to one.

Natural diarization windows are 2–120 seconds, with at most two longer VoxConverse windows per recording, no more than 1,200 seconds. Long windows are chosen by source speaker-count support then seeded hash. No synthetic compression creates higher counts. Smart Turn clips retain their supplied duration within the accepted 0.2–30-second range.

Source attribution and modifications

Source	Immutable revision	Attribution
Smart Turn v3.2 train	f8c95d284595defb6b88a715c356cbc5c31b7ad1	Pipecat team, Liva AI, Midcentury, MundoAI, and contributors listed in the preserved upstream card
Smart Turn v3.2 test	dfc179241fee65b0efe81ca37786f41ce26f38ec	Same contributors; external diagnostic only
AMI SDM	8cdaae2eaf968f3b000b6eb1204ab9b8db006ed0	AMI Consortium; McCowan, Carletta, Kraaij et al., The AMI Meeting Corpus (2005); diarizers preprocessing contributors
VoxConverse	3acfa1b45ca4b7419aee999d67d94c617f9c9d47	Chung, Huh, Nagrani, Afouras, Zisserman, Spot the conversation: speaker diarisation in the wild (Interspeech 2020); original creators and diarizers preprocessing contributors

VoxConverse timestamps use corrected upstream v0.3 RTTM at commit 24bf60be297701cd7e4ef18550c6d390c1b87365. AMI participant linkage uses the official manual 1.6.2 annotation archive. Exact source asset hashes, byte receipts, source cards, contribution links, and licenses are retained in provenance/.

Derived changes: deterministic selection and split assignment; physical cropping; mono downmixing; PCM16 FLAC encoding; categorical runtime questions and labels. This is a derivative compilation, without endorsement by source authors. All newly included sources are distributed as CC-BY-4.0 according to the pinned source cards and upstream notices. Preserve source attribution when redistributing. Original-task replay and retention corpora, which have separate licensing requirements, are not part of this publication.

Splits and limitations

AMI sessions connected by any shared global participant are assigned as one component. Every component touching prior Audio Jev participants is excluded from this new corpus. Remaining components use seeded 70/10/10/10 quotas. VoxConverse official test stays test; official development recordings are assigned whole to train/dev/calibration. VoxConverse has no global cross-video speaker identity guarantee.

Smart Turn natural training uses human_convcollector_1; multilingual synthetic training uses chirp3_2. Natural development uses midcentury_1, calibration human_5, and test liva_1. Prior row identities and hashes are excluded. These collections were encountered in earlier project evaluation: this is a fresh held-out identity/window partition within known source corpora, not a novel-corpus claim. Individual Smart Turn speaker/session identities are unknown, so collection separation does not establish person separation.

Strict splits reject exact encoded audio and decoded canonical PCM overlap; source lineage is checked again by the Parquet SDK. Edited near-duplicates and unknown cross-recording identities remain possible. Windows overlap within recordings and must not be treated as independent people or meetings. Use recording/source-component confidence intervals. Report natural and synthetic subsets separately.

Count/activity/duration strata are selected for coverage, not population prevalence. Calibration applies to this selected distribution; unknown assistant deployment prevalence requires separate validation. Some source annotations extend beyond the captured audio. Original timestamps are preserved; counts/activity intersect only captured windows, and future labels require two seconds of captured look-ahead. The source quality audit records wholly post-capture and crossing-end spans without asserting their cause. Timestamp boundary error, incidental/background voices, domain shift from meetings/broadcast audio, and short audible speaker evidence can affect targets. Backchannel and mid-thought-pause gold labels are not supplied by the diarization proxy.

Reading the normalized Parquet tables

The audio table stores each distinct recording/window blob once. decisions references it by recording_id; audio_id identifies exact observed frame bounds. Labels, targets, context, and provenance are JSON strings in the decision table. A generic Hugging Face viewer does not join these tables automatically.

from huggingface_hub import snapshot_download
from jev_audio_eval.datasets.parquet import ParquetCorpus

root = snapshot_download("sammyliu/audio-jev-rlcd-v1", repo_type="dataset")
corpus = ParquetCorpus(root)
with corpus.samples("train", audio_cache_dir="./audio-jev-cache") as samples:
    sample = samples[127]  # Audio materializes only when requested.
    for row in samples.iter_metadata(batch_size=64):
        task = row["task"]  # Metadata scan reads no waveform.

Install the SDK from samliu/audio-jev. The external diagnostic opens with ParquetCorpus(root + "/external"). Each split includes a checksummed manifest and SQLite row-group index. Default row-group caching is bounded; bulk audio and retained feature caches can remain on NAS.

For Hugging Face streaming, load_dataset("sammyliu/audio-jev-rlcd-v1", "decisions", split="train", streaming=True) scans decisions without loading audio. Join the separate audio config on recording_id, or use the SDK above. Avoid copying targets or provenance into model-visible questions.

Reproducibility

Source plan SHA-256: 365c4d3570af5ca099afc8fc49a72294cf0cb0ae2c0766094abfd11687e7b6fb. Partition lock SHA-256: 90346c4ecf9b433f899cfd86ad07bf4bdd3d34da4c27d2470ca5d1ee1dbd7ace. Build receipt SHA-256: 35267d7f8f9749ec5218b62ed5caf614341ed7ea8dd1480c2683b276b4d0e6e2.

See provenance/build.json for exact task/count coverage, duration/token estimates, content deduplication removals, and split checksums; provenance/partitions.json contains the source-only split lock. Selection uses no model scores. No claim of trained model quality is made by this dataset card.

Corpus SDK implementation: samliu/audio-jev@1551fc6.