Humaneness intense emotion and talking-style subsets
Gated research collection. No new collection license is granted. This does not remove original source licenses, attribution or restrictions. Access approval is not a declaration that the material is unrestricted or public-domain. Existing notices are retained in source metadata.
Publication status: packed payload upload completed. Do not treat every candidate as an acoustically verified example.
The screened pool contains 45,112 unique clips / 272.3126 unique audio hours, from accepted synthetic acting and annotated real-speech training targets. It has 40 emotion heads, 15 literal talking styles, eight high/low dimensions and two formal/casual contrasts. All five languages are represented globally, not in every individual bucket. Cross-bucket overlap is intentional; summed membership hours are not unique hours.
Important taxonomy correction
vn_S_MONO_reg means Monologue Style, not monotone. vn_VOLT_reg means Volatility, not loudness. vn_S_TECH_reg is Teacher/Didactic Style. Earlier interpretations and their incompatible audio judgments are retained locally as historical audit evidence, not treated as valid judgments for the corrected categories. No new loudness head is invented. Canonical definitions come from the frozen procedural caption taxonomy.
Initial score-filtered selection and verification
Candidates must be in a language-specific extreme 1% tail, meet explicit provisional absolute score cutoffs, CE≥4.5 and Production Quality≥6, and retain the original split/deduplication rules. The activation proxy for anger uses arousal plus ranting OR volatility; the latter is not a loudness measurement. No cutoff is weakened simply to fill an hour target. Empty/small buckets remain explicit.
An independent Gemini 3.8 Flash reviewer receives audio and acoustic hypotheses, but no transcript, source instructions or EARS scores. Passing a bucket requires sustained audible intensity, usable recording quality, one speaker and valid timestamp evidence. Machine scores/reviews are not human ground truth. Invalid judgments never certify a clip. Corrected taxonomy labels are reviewed again, not silently relabeled.
index/screened-memberships.parquet lists model-screened candidates. index/audio-verified-memberships.parquet contains only individually passing UID/bucket pairs when published. Training adapters use the latter, not all raw candidates. Sparse/unsupported buckets cannot honestly produce a trained adapter.
Packed format
Read data/*.tar without expanding the corpus into loose files. Original audio, MOSS-v2 arrays, annotations, captions, timing and embeddings remain in each UID group. Twelve codec codebook indices occur per 80 ms frame: 12.5 frames/s, or 150 indices/s, not 12 indices/s. MP3 playback and codec targets are not regenerated. The public index records byte offsets and hashes. Available reference-code crops are included as UID.reference.npy, with 63–187 frames (5.04–14.96 s); their original waveforms are not duplicated. Reference matching remains inherited, not a new claim of verified real-speaker identity.
from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq
p = hf_hub_download("laion/Humaneness-Intense-Emotion-Style-Subsets", "index/screened-memberships.parquet", repo_type="dataset")
subset = pq.read_table(p, filters=[("language", "=", "de"), ("bucket", "=", "emo_Anger")])
Provenance and scope
Synthetic source: LAION acting challenges, including multilingual Gemini and additional ElevenLabs voices. Real sources: EN/DE RealSpeech and the five-language annotated collection. Original source-specific rights remain applicable; this repository does not replace or erase them. Local machine paths are not required to use the public members.
Bucket inventory
| Bucket | Correct acoustic meaning | Screened memberships | Membership hours |
|---|---|---|---|
| emo_Affection | Intense Affection | 1045 | 8.282 |
| emo_Amusement | Intense Amusement | 1554 | 9.998 |
| emo_Anger | Intense Anger | 768 | 4.799 |
| emo_Astonishment_Surprise | Intense Astonishment Surprise | 1517 | 9.998 |
| emo_Awe | Intense Awe | 791 | 5.892 |
| emo_Bitterness | Intense Bitterness | 510 | 3.394 |
| emo_Concentration | Intense Concentration | 977 | 5.395 |
| emo_Confusion | Intense Confusion | 239 | 1.291 |
| emo_Contemplation | Intense Contemplation | 1144 | 9.998 |
| emo_Contempt | Intense Contempt | 423 | 2.791 |
| emo_Contentment | Intense Contentment | 649 | 4.323 |
| emo_Disappointment | Intense Disappointment | 1223 | 9.995 |
| emo_Disgust | Intense Disgust | 336 | 2.116 |
| emo_Distress | Intense Distress | 1250 | 9.995 |
| emo_Doubt | Intense Doubt | 4 | 0.032 |
| emo_Elation | Intense Elation | 174 | 0.938 |
| emo_Embarrassment | Intense Embarrassment | 0 | 0.000 |
| emo_Emotional_Numbness | Intense Emotional Numbness | 0 | 0.000 |
| emo_Fatigue_Exhaustion | Intense Fatigue Exhaustion | 1087 | 9.990 |
| emo_Fear | Intense Fear | 1432 | 9.992 |
| emo_Helplessness | Intense Helplessness | 1209 | 9.996 |
| emo_Hope_Enthusiasm_Optimism | Intense Hope Enthusiasm Optimism | 1299 | 8.446 |
| emo_Impatience_and_Irritability | Intense Impatience and Irritability | 1553 | 9.998 |
| emo_Infatuation | Intense Infatuation | 31 | 0.220 |
| emo_Interest | Intense Interest | 1809 | 9.998 |
| emo_Intoxication_Altered_States_of_Consciousness | Intense Intoxication Altered States of Consciousness | 0 | 0.000 |
| emo_Jealousy_and_Envy | Intense Jealousy and Envy | 0 | 0.000 |
| emo_Longing | Intense Longing | 523 | 4.479 |
| emo_Malevolence_Malice | Intense Malevolence Malice | 8 | 0.044 |
| emo_Pain | Intense Pain | 742 | 6.379 |
| emo_Pleasure_Ecstasy | Intense Pleasure Ecstasy | 0 | 0.000 |
| emo_Pride | Intense Pride | 43 | 0.213 |
| emo_Relief | Intense Relief | 668 | 5.055 |
| emo_Sadness | Intense Sadness | 1021 | 9.136 |
| emo_Sexual_Lust | Intense Sexual Lust | 7 | 0.047 |
| emo_Shame | Intense Shame | 18 | 0.108 |
| emo_Sourness | Intense Sourness | 449 | 3.210 |
| emo_Teasing | Intense Teasing | 642 | 4.464 |
| emo_Thankfulness_Gratitude | Intense Thankfulness Gratitude | 1219 | 8.224 |
| emo_Triumph | Intense Triumph | 57 | 0.279 |
| vn_S_ASMR_reg | Strong ASMR Style | 99 | 0.847 |
| vn_S_AUTH_reg | Strong Authoritative Style | 2157 | 9.664 |
| vn_S_CART_reg | Strong Cartoonish Style | 0 | 0.000 |
| vn_S_CASU_reg | Strong Casual Style | 2518 | 9.999 |
| vn_S_CONV_reg | Strong Conversational Style | 2293 | 9.999 |
| vn_S_DRAM_reg | Strong Dramatic Style | 1398 | 9.997 |
| vn_S_FORM_reg | Strong Formal Style | 919 | 3.697 |
| vn_S_MONO_reg | Strong introspective monologue delivery, not monotone speech | 1368 | 9.993 |
| vn_S_NARR_reg | Strong Narrator Style | 1810 | 8.453 |
| vn_S_NEWS_reg | Strong Newsreader Style | 0 | 0.000 |
| vn_S_PLAY_reg | Strong Playful Style | 2042 | 9.152 |
| vn_S_RANT_reg | Strong Ranting/Angry Style | 40 | 0.164 |
| vn_S_STRY_reg | Strong Storytelling Style | 513 | 2.858 |
| vn_S_TECH_reg | Strong teacher-like, didactic delivery | 14 | 0.053 |
| vn_S_WHIS_reg | Strong Whisper-Talk Style | 16 | 0.149 |
| vn_VULN_reg_high | emotionally vulnerable and exposed | 1162 | 9.992 |
| vn_VULN_reg_low | emotionally detached and unexposed | 2506 | 9.226 |
| vn_STNC_reg_high | dominant and assertive | 2219 | 9.999 |
| vn_STNC_reg_low | submissive and deferential | 1062 | 9.991 |
| vn_AROU_reg_high | highly aroused and activated | 1664 | 9.996 |
| vn_AROU_reg_low | very low arousal, calm or sleepy | 125 | 1.004 |
| vn_VOLT_reg_high | Emotionally volatile delivery with rapid or abrupt changes | 1371 | 9.987 |
| vn_VOLT_reg_low | Emotionally steady, stable delivery without abrupt changes | 1378 | 2.937 |
| vn_VALN_reg_high | strongly positive vocal affect | 2289 | 9.999 |
| vn_VALN_reg_low | strongly negative vocal affect | 1215 | 9.995 |
| vn_WARM_reg_high | very warm vocal delivery | 1156 | 6.638 |
| vn_WARM_reg_low | very cold vocal delivery | 0 | 0.000 |
| vn_TENS_reg_high | very tense and strained | 1503 | 9.995 |
| vn_TENS_reg_low | very relaxed and unstrained | 2 | 0.005 |
| vn_TEMP_reg_high | very fast speaking tempo | 2332 | 7.340 |
| vn_TEMP_reg_low | very slow speaking tempo | 26 | 0.201 |
| casual_not_formal | Very informal/casual, not formal | 2518 | 9.999 |
| formal_not_casual | Very formal, not casual | 917 | 3.690 |
Adapter comparisons · Parent rank-4 model bundle.
Audio-verified tail extension for initially sparse buckets
The initial collection above used provisional fixed raw EARS cutoffs. Several heads never reached those cutoffs, so a new discovery tier uses the strongest1% tail per language/head without that provisional absolute floor. This tier is not automatically certified. The same strict independent audio-only acceptance still requires intensity≥3/4,confidence≥3/4,sustained fraction≥0.5,correct language,one speaker,quality≥3/4 and enjoyment≥2/4. Source CE/PQ and real-speech transcript-consistency gates are unchanged. No weak clip is included merely to fill a bucket.
Additional exact source audio,all12 MOSS-v2 codebooks,captions and metadata are under data/gap-repair/; byte-offset indices are under index/gap-repair/. Existing initial-pool audio is reused rather than duplicated. index/audio-verified-gap-repair.parquet lists passing UID/bucket pairs for the supplemental adapters, including reused initial positives. For a full verified membership view, concatenate this file with the original index/audio-verified-memberships.parquet and deduplicate (uid,bucket). Read both the original and gap-repair packed-member indices when resolving UIDs. GAP_DISCOVERY.json records the per-head/language calibration. The initial and supplemental cohorts remain distinguishable.
Access remains manually gated, with no new collection license. Original notices remain applicable.
- Downloads last month
- 54