IEMOCAP — full release, utterance level

All 10,039 segmented utterances of the USC-SAIL IEMOCAP corpus with 16 kHz audio, transcripts, consensus emotion labels, consensus VAD ratings, per-annotator raw labels, and annotator-agreement statistics.

This is a private mirror. IEMOCAP is distributed under the USC SAIL academic license, which requires an individually signed agreement and does not permit redistribution. Do not make this repository public. Anyone needing the data should obtain it from sail.usc.edu/iemocap.

Loading

from datasets import load_dataset, concatenate_datasets

ds = load_dataset("cairocode/iemocap")          # 5 splits, one per session
ds["session1"][0]["audio"]                      # {'array': ..., 'sampling_rate': 16000}

# Leave-one-session-out: hold out session 5
train = concatenate_datasets([ds[f"session{i}"] for i in (1, 2, 3, 4)])
test = ds["session5"]

# Standard 4-class benchmark subset (5,531 utterances)
four = train.filter(lambda r: r["emotion_4class"] is not None)

Splits are the five recording sessions, each with a disjoint pair of actors, so session-wise splitting is automatically speaker-independent. Concatenate all five for the full corpus.

Fields

Identification

Field	Type	Description
utterance_id	str	e.g. Ses01F_impro01_F000
session	int	Recording session, 1–5
dialog_id	str	Parent dialogue, e.g. Ses01F_impro01
script_type	str	improvisation (4,784) or script (5,255)
speaker_id	str	One of 10 actors, e.g. Ses01F
speaker_gender	str	female (4,800) / male (5,239)
marker_actor_gender	str	Which actor wore the MoCap markers in this dialogue
start_time, end_time, duration	float	Seconds, relative to the parent dialogue

Audio and text

Field	Type	Description
audio	Audio	16 kHz mono, PCM-16, embedded. Native rate — nothing was resampled
transcript	str	Manual transcription. Present for all 10,039 rows

Consensus labels

Field	Type	Description
emotion	str	Majority-vote code, or xxx when no majority exists
emotion_full	str	Same, spelled out (undecided for xxx)
emotion_4class	str | None	angry / happy / neutral / sad, with exc merged into happy. None for the other 4,508 rows
valence, activation, dominance	float	Corpus-provided mean over the external dimensional annotators, on a 1–5 scale. Self-evaluations are excluded, per the corpus README

emotion is the corpus's own ground-truth field, not recomputed here.

Agreement

Field	Type	Description
n_annotators_categorical	int	Always 3 external annotators
n_annotators_attribute	int	Dimensional annotators: 2 (8,666), 3 (1,279), 4 (58), 1 (36)
emotion_votes	JSON str	Category → vote count, e.g. {"fru": 2, "neu": 1}
majority_votes	int	Votes for the most-chosen category
agreement_ratio	float	majority_votes / 3. Corpus mean 0.673
unanimous	bool	All 3 annotators gave the same primary label (2,132 rows, 21.2%)
has_majority	bool	emotion != "xxx" (7,532 rows, 75.0%)
valence_std, activation_std, dominance_std	float	Sample standard deviation across external dimensional annotators

Annotators could assign more than one category to an utterance, so vote counts in emotion_votes may sum to more than 3. agreement_ratio and unanimous are computed over each annotator's first (primary) label.

Corpus-level Fleiss' κ = 0.276 on primary labels (3 raters, 10 categories, N = 10,039) — low, which is characteristic of the corpus and the reason a quarter of utterances have no majority label.

Per-annotator raw labels

Field	Type	Description
annotators_categorical	JSON str	{"E2": ["neu"], "E3": ["neu"], "E4": ["neu", "ang"]}
annotators_attribute	JSON str	{"E3": [3, 2, 2], "E4": [2, 3, 3]} — [valence, activation, dominance], 1–5
annotators_categorical_comments	JSON str	Free-text annotator notes, where given
annotators_attribute_comments	JSON str	As above, for dimensional ratings
self_categorical	JSON str | None	The performing actor's own categorical rating
self_attribute	JSON str | None	The actor's own [val, act, dom]

Annotator codes are the corpus's own (E1–E6 external; F*/M* the actors themselves). These are not stable identities across sessions: E2 in session 1 is not necessarily E2 in session 3. Do not model annotator identity across sessions.

Self-evaluations cover only 2,409 of 10,039 utterances, and session 4 has none at all — these columns are None elsewhere. They are excluded from the consensus VAD.

JSON-encoded columns are strings; json.loads them.

Label distribution

Code	Emotion	Count	Share
xxx	undecided	2,507	25.0%
fru	frustrated	1,849	18.4%
neu	neutral	1,708	17.0%
ang	angry	1,103	11.0%
sad	sad	1,084	10.8%
exc	excited	1,041	10.4%
hap	happy	595	5.9%
sur	surprised	107	1.1%
fea	fearful	40	0.4%
oth	other	3	0.0%
dis	disgusted	2	0.0%

4-class subset: neutral 1,708 · happy 1,636 (595 hap + 1,041 exc) · angry 1,103 · sad 1,084 = 5,531.

Rows per session: 1,819 / 1,811 / 2,136 / 2,103 / 2,170. Total speech: 12.44 hours.

Provenance

Built from IEMOCAP_full_release by parsing the aggregate SessionX/dialog/EmoEvaluation/*.txt files (151 dialogues), joined to SessionX/dialog/transcriptions/ and SessionX/sentences/wav/. Every utterance matched a wav file and a transcript; no rows were dropped or imputed.

The aggregate EmoEvaluation files are the authoritative source. The EmoEvaluation/Categorical/, Attribute/ and Self-evaluation/ subdirectories duplicate the same annotations but encode the numeric scales on an inconsistent polarity — the same annotator, utterance and free-text comment appears there with different numbers. Those subdirectories were deliberately not used.

Not included

Dialogue-level audio and video, MoCap (face, head, hand), forced alignments (ForcedAlignment/*.wdseg|phseg|syseg), and the .anvil annotation sources. All remain in the local IEMOCAP_full_release tree and can be added on request.

Citation

Any published work using IEMOCAP must cite:

@article{busso2008iemocap,
  title={IEMOCAP: Interactive emotional dyadic motion capture database},
  author={Busso, Carlos and Bulut, Murtaza and Lee, Chi-Chun and Kazemzadeh, Abe and
          Mower, Emily and Kim, Samuel and Chang, Jeannette N. and Lee, Sungbok and
          Narayanan, Shrikanth S.},
  journal={Language Resources and Evaluation},
  volume={42}, number={4}, pages={335--359}, year={2008},
  publisher={Springer}
}