laion/whisper-gemini-voice-annotations

LAION Gemini voice annotations: emotion, delivery, quality and vocal events

Christoph Schuhmann / LAION · 2026-10-07 · CC BY 4.0 for the new annotation files and included source audio. This is the validated annotation pool used for the two-epoch Humaneness Ears Base and Medium release. The dataset holds 65,163 unique SHA256 audio identities, with train/validation/test counts of 58,380/3,239/3,544. 1,036 blocked or schema-invalid Gemini responses were excluded; they are not negative examples or zero-valued emotions.

Each valid example has a Gemini 3.8 Flash response with transcription, whole-audio and director captions, 40 simultaneous EmoNet emotion intensities, 57 VoiceNet dimensions, genuineness, vocal-burst blend, a burst count and labeled events with explicit start and end seconds. Additional teacher outputs include Empathic Insight Voice Plus, AudioBox, DNSMOS, VoiceCLAP and Orange timbre/identity embeddings. These are model annotations, not human ground truth. The annotation values should be reviewed before applications requiring human-label validity.

Sources and audio rights

Source family	Valid unique records in this release	Audio in TAR?	Attribution and rights
S1 Scaling Ladder	42,701	Only records individually marked CC BY 4.0	S1 is mixed-license as a whole; each published audio record carries its original per-sample license.
LAION’s Got Talent raw	10,917	Yes	LAION publisher has directed this derivative release under CC BY 4.0. Each sample retains its voice, emotion, language, source archive and source ID.
Balanced 40×3k	3,979	No	Only our newly produced annotations and embedding outputs are distributed. The upstream repo has no declared audio license in its Hub metadata.
Voice Annotation POC	7,566	No	Only our newly produced annotations and embedding outputs are distributed. The upstream repo has no declared audio license in its Hub metadata.

This package contains 52,036 exact MP3s and 13,127 annotation-only records. A top-level CC BY 4.0 license does not relicense audio retained only in upstream repositories. The source URL, revision or original source member and audio SHA are provided for tracing an annotation-only record. No unverified mitermix audio is copied. The entire data package is deduplicated by compressed-audio SHA256; transcoded duplicates with different bytes are not guaranteed to be removed.

Files and schema

The 516 WebDataset TAR shards under data/<source-family>/<split>-NNNN.tar hold up to 128 examples each. A key is the full lowercase MP3 SHA256:

• <sha>.json: source and split provenance, license flag, full Gemini annotation, 192 teacher scalar outputs in native units, score provenance, CPS, exact burst spans and masks.
• <sha>.npz: float16 arrays voiceclap (768D), bud_e_pooled (3072D), timbre (128D) and identity (250D). Orange vectors come from the same selected waveform; speaker_target_valid indicates whether the supervised single-speaker mask was active.
• <sha>.mp3: present only when audio_in_shard is true. It is byte-identical to the selected source and validated by SHA256.

manifest.jsonl lists every key, source family, split, shard path, duration and audio-presence flag. DATASET_READY.json gives audited counts. example.mp3 and example.json provide a directly playable full example. For source audio absent from a shard, use the provided source repository under its own terms. Audio is never silently fabricated or replaced with a clean speaker reference.

A complete annotated example

Play or download the example MP3 · Inspect the full JSON annotation · Open the interactive listening atlas

This example is 10.22 seconds long, SHA256 00003c59c3999f3d6a0ebbe97bce67053785d09a0e67988ded26bf19035f1938, from vc_sidon. Its Gemini response contains the transcript, whole-audio caption, director caption, all 40/57 scores and the following event spans:

Event	Start (s)	End (s)	Fine vocal-event label
1	0.00	1.95	Heavy Breathing
2	2.05	4.50	Exasperated Sigh
3	6.80	9.80	Heavy Breathing

The linked JSON contains the complete 192-score sidecar, CPS, event captions and teacher provenance. The corresponding vector .npz is in the manifest-listed shard.

Annotation and training decisions

The original detailed annotation prompt and the released preparation/validation code specify the exact schema. Gemini was passed the complete selected audio once per clip. A schema-valid response takes precedence over an invalid attempt. Unusable responses are excluded. Prompt hash, model ID and batch tag are retained per example.

The 40 EmoNet scores are simultaneous 0–4 intensities, following EmoNet. The 57 VoiceNet dimensions mostly use 0–6; BKGN uses 0–4 and EXPL 0–2. See VoiceNet taxonomy and benchmark and the VoiceNet model repository. Genuineness is 0–6 (head); burst blend is 0–10 (head). Every burst preserves its original fine label, coarse group, caption and unclipped start/end seconds. The released student maps these into its 53-class vocabulary while retaining the original 78-label annotation taxonomy in the data.

Other packages used: Empathic Insight Voice Plus, facebook/audiobox-aesthetics, Microsoft DNSMOS, Orange Speaker-wavLM-tbr and Orange Speaker-wavLM-id. The dataset keeps each scale and construct separate: AudioBox content enjoyment is not VoiceNet aesthetics; DNSMOS BAK is not the Empathic background-noise score. Timed vocal events are separate from clip-level burst count. Literal characters per second include transcript spaces and punctuation, divided by whole clip duration, after bracketed nonlexical placeholders are removed.

Train/validation/test labels are those of the Gemini fine-tune, grouped by known speaker or synthetic audio family where available. They do not prove these clips were unseen during the models' earlier S1–S10 training. Refer to the independent benchmark report for measured generalization and public human-label comparisons.

Citation and use

Attribute the new annotations and package to Christoph Schuhmann / LAION, 2026, under CC BY 4.0, and retain the per-record upstream attribution. For annotation-only rows, consult the upstream source before acquiring or redistributing audio. Model outputs are continuous estimates, not calibrated probabilities or identity guarantees.