This dataset contains synthetic speech in the voices of real, identifiable people. One utterance inside each counterfactual clip was re-synthesised in another speaker's voice using voice cloning.

The source corpora (AMI, ICSI, VoxMM) carry licences or research-use conditions for the original recordings, but those conditions do not constitute consent to generate novel utterances in an individual's voice, and no task-specific voice-cloning consent was obtained from the source speakers. We also do not treat pseudonymous speaker ids as full anonymity, because a recognisable voice itself carries identity. For that reason the synthetic audio is released as a controlled research resource rather than an unrestricted download.

Data Use Agreement

By requesting access you agree that you will:

Use the synthetic audio only for non-commercial research on speaker-aware
speech understanding and reasoning.
Not redistribute the audio or any derivative of it, in whole or in part.
Not attempt to re-identify the source speakers.
Not use the audio for impersonation, deceptive or misleading content, fraud,
harassment, or any use intended to harm or misrepresent an individual.
Not use it for commercial voice replication, and not deploy cloned voices in
interactive or production systems.
Treat every synthetic utterance as artificially generated. These are not
statements actually made by the corresponding speakers and must not be presented
as representing their views, intentions, beliefs, or endorsements.

Access is revocable if these conditions are violated.

Takedown

Rights holders, individuals whose voice appears in the data, and their authorised representatives may request removal. Identify the source recording or give enough information to locate the affected samples; valid requests are removed from subsequent distributions. Takedown records are kept in the dataset documentation without recording unnecessary personal information about the requester.

HEAR

🎉 EMNLP 2026 main conference 🎉

🌐 Project page · 💻 Code · 🤗 Dataset · 🧠 Model

Hierarchical Evaluation of Attribution and Reasoning, a benchmark for speaker-attributed understanding of multi-party speech.

Most speech benchmarks can be solved by transcribing the audio and reading the text. HEAR cannot. Every question asks something about who is speaking, not only what is said, and roughly half the benchmark comes in counterfactual pairs: the same clip, the same question, the same five options, but one utterance has been re-recorded in a different speaker's voice, which flips the correct answer. A model that only reads the transcript answers both members of a pair identically and scores zero on the pair.

from datasets import load_dataset

hear = load_dataset("PleasedPenguin/HEAR", split="test")
row = next(r for r in hear if r["subtaxonomy"] == "VC")

row["audio"]       # the prompt audio, exactly what the model should hear
row["question"]    # "How many distinct speakers are present in this audio clip?"
row["options"]     # ['5', '2', '6', '3', '4']
row["answer_idx"]  # 3   -> option "3", letter "D"
row["answer"]      # "(D) 3"

Why it is hard

Multi-party, real recordings. 887 clips from meetings and YouTube conversations, 4.05 speakers per clip on average, 22.8 hours in total.
Overlapping speech. Every dimension except Voice Cardinality ships an overlap subset, where answering requires attending to a region with simultaneous speech.
Counterfactual voice swaps. 580 pairs where one utterance was re-voiced. The transcript is unchanged; the answer is not.
Voices as answer options. In Content-to-Voice Attribution the five options are voice samples, not text.

Taxonomy

Three axes → eight sub-dimensions. Descriptions follow the paper.

Discrimination, tell voices apart and locate them in time

	Sub-dimension	What it asks	n
`VC`	Voice Cardinality	How many unique speakers are present in the clip	190
`VL`	Voice Localization	Given a reference voice, which time range does that speaker appear in, or not appear in	237
`VCD`	Voice Change Detection	Which time range contains a speaker transition (non-overlap) or simultaneous speech (overlap)	200

Attribution, bind content to identity, in both directions

	Sub-dimension	What it asks	n
`CVA`	Content-to-Voice Attribution	Given an utterance as text, pick the speaker from five candidate voices	307
`VCA`	Voice-to-Content Attribution	Given a reference voice, pick which of five texts that voice spoke	301

Reasoning, track "who said what" across the conversation

	Sub-dimension	What it asks	n
`IR`	Identity Reasoning	Given an anchor utterance, find another utterance by the same speaker	370
`QR`	Quantitative Reasoning	Aggregate per-speaker statistics, count speakers who said a phrase (`QR1`), or rank speakers by total speaking time (`QR2`)	560
`TR`	Temporal Reasoning	Order utterances by speaker, a speaker's first/last utterance (`TR1_first`, `TR1_last`), or what they said after an anchor by someone else (`TR2`)	230

`task` and `overlap`

task_name names the sub-task inside a sub-dimension, and is none where a sub-dimension has only one form (VC, CVA, VCA, IR):

change · overlap (VCD), appear · notappear (VL), QR1 · QR2, TR1_first · TR1_last · TR2

overlap says whether answering requires attending to overlapping speech. It is annotated by human reviewers for every dimension except Voice Cardinality, where it is null. For VCD the two task_name values are the two regimes, so change → nonoverlap and overlap → overlap.

The counterfactual pairs

variant is original or hallucinated for the 1,160 Reasoning rows, and none elsewhere. Twins share everything but the voice of one utterance:

reasoning = hear.filter(lambda r: r["variant"] != "none")

# a pair, recovered from the id
base = "Bmr019_spk5_len111s_31b1_IR_overlap_fe008_76.3_mn017"
orig = hear.filter(lambda r: r["question_id"] == base + "_original")[0]
hall = hear.filter(lambda r: r["question_id"] == base + "_hallucinated")[0]

orig["question"] == hall["question"]      # True
orig["options"]  == hall["options"]       # True
orig["answer_idx"], hall["answer_idx"]    # 0, 2   <- the voice swap moved the answer

There are 580 pairs, all complete, every _original has its _hallucinated, and no row is an orphan.

Columns

column	type	notes
`question_id`	string	primary key; a Reasoning id ends in `_original` / `_hallucinated`
`clip_id`	string	887 clips; several questions share one clip
`source_corpus`	string	`ami` · `icsi` · `voxmm`
`taxonomy`	string	`Discrimination` · `Attribution` · `Reasoning`
`subtaxonomy`	string	the eight above
`task_name`	string	sub-task, `none` where there is only one
`overlap`	string	`nonoverlap` · `overlap` · `null` (Voice Cardinality only)
`variant`	string	`original` · `hallucinated` · `null`
`question`	string	the question alone
`options`	list[string]	five options, already shuffled, this order is authoritative
`answer_idx`	int32	0-based index into `options`
`answer`	string	`"(D) 3"`, letter and text together
`audio`	Audio 16 kHz	the prompt, the clip plus whatever reference audio the task needs
`option_audio_A` … `option_audio_E`	Audio 16 kHz	the five candidate voices, one column each so the viewer plays them; set for Content-to-Voice Attribution, `null` elsewhere
`answer_audio`	Audio 16 kHz	the correct voice, for the same rows, `null` wherever the answer is text

Three prompt layouts, all pre-assembled in `audio`

layout	`audio` contains	sub-dimensions
main only	the clip	VC · VCD · IR · QR · TR
main + reference	clip, then the target speaker's voice	VL · VCA
main + five voices	clip, then "A" ⟨voice⟩ … "E" ⟨voice⟩	CVA

You never have to assemble anything: audio is the exact waveform used in the paper. The per-option voice columns are provided separately so you can rebuild a different prompt format if you want.

Scoring

Plain accuracy for Discrimination and Attribution. Reasoning is scored twice:

# per-item accuracy, counting only the ORIGINAL member of each pair
# pair-strict: a pair scores 1 only if BOTH members are correct

The paper's headline metric, Ver-A, treats a counterfactual pair as one unit:

units  =  1,235 non-Reasoning items  +  580 pairs  =  1,815

Pair-strict scoring is what separates voice-grounded models from transcript-reading ones. Strong text-centric models lose 30–50 points when they move from per-item to pair-strict.

Source data

Clips are drawn from three corpora with human-produced transcripts, segmented to 30–150 s and filtered to multi-party regions.

corpus	clips	source recordings	speakers	mean length	hours
VoxMM (YouTube)	360	34	245	87.8 s	8.78
AMI (meetings)	241	23	24	97.4 s	6.52
ICSI (meetings)	286	6	21	94.9 s	7.54
total	887	63	290	92.7 s	22.84

Every question was reviewed by human annotators for solvability, a unique correct answer, voice-identity preservation in the swapped clips, and transcript–option consistency.

Note on counts. The paper reports 2,575 questions over 990 clips. This release is the camera-ready revision: Ego4D was removed for licensing reasons (159 questions, 103 clips) and 21 further items were dropped so that every counterfactual pair is complete and every row's audio matches its metadata. Hence 2,395 questions over 887 clips.

Also in this repo: CASH-60K, the training corpus

CASH-60K/ holds the corpus A2R was trained on, 59,762 queries built from the VoxMM train split, with hard negatives made the same way HEAR's counterfactuals are: one utterance re-recorded in another speaker's voice, so the words are unchanged and the answer moves.

It is kept out of the dataset viewer on purpose, the viewer is for the benchmark. Because it carries columns the benchmark does not, load it as parquet rather than by repo id, otherwise datasets tries to cast it into the benchmark's schema:

from datasets import load_dataset

cash = load_dataset(
    "parquet",
    data_files="hf://datasets/PleasedPenguin/HEAR/CASH-60K/train-*.parquet",
    split="train",
)

Same column names as the benchmark, plus two, and with no answer_audio:

column	notes
`variant`	`original` · `hard_negative`
`hn_version`	which swap recipe produced a hard negative (`v0`–`v7`, `real_v5`); `null` on originals
`pair_id`	groups an original with its 1–3 hard negatives

	queries
original	25,943
hard negative	33,819
total	59,762

Built from 5,242 source clips plus 14,180 voice-swapped variants (19,422 clips total). Every synthesized swap was screened by ASR word-error-rate against the intended transcript, so only the voice moved. CASH draws from VoxMM train; HEAR draws from VoxMM test plus AMI and ICSI, no clip overlap.

Also in this repo: the three OOD benchmarks

ood_benchmark/ holds the three held-out benchmarks A2R is evaluated on zero-shot. None of them is drawn from the corpora CASH-60K is built on, and in all three the answer turns on voice identity rather than on the transcript.

file	task	questions	clips	hours	mean clip	options
`wdyl-*.parquet`	a two-speaker dialogue, then a first-person question asked in one of the two voices	793	793	2.36	10.7 s	3
`gaokao-*.parquet`	a listening-exam conversation, then a first-person question in one speaker's voice	93	93	0.52	20.0 s	2
`fts-*.parquet`	Find the Spy: 3-6 players introduce themselves, then give clues in a reshuffled order without names	400	400	3.67	33.1 s	3-6

1,286 questions and 6.55 hours in total, all multiple choice with one gold option. Columns are question_id, benchmark, audio, question, options, answer_idx, answer, instruction. Audio decodes to 16 kHz mono. instruction carries the task description and the options and stops there, so you are free to append whatever output format your model expects; the paper's own numbers were produced by appending a transcribe-then-reason instruction to it.

from datasets import load_dataset

fts = load_dataset(
    "parquet",
    data_files="hf://datasets/PleasedPenguin/HEAR/ood_benchmark/fts-00000-of-00001.parquet",
    split="train",
)

Ethical considerations

This resource contains synthetic speech in the voices of real people. Access is gated for that reason; the terms are in the access form and summarised here.

Dual use. Voice cloning and speaker attribution can be misused for impersonation, deceptive audio, fraud, harassment, or speaker re-identification. We therefore treat the synthetic speech as a controlled research resource, not an unrestricted public download.

Voice, likeness, consent. The source corpora license the original recordings. That is not consent to generate new utterances in someone's voice, and no task-specific voice-cloning consent was obtained from the source speakers. A recognisable voice carries identity on its own, so pseudonymous speaker ids are not treated as anonymity.

What is open and what is gated. The code, the evaluation protocol, the annotations and the aggregate statistics are public. The synthetic waveforms are available only through this gated repository, to approved non-commercial research, under the Data Use Agreement.

Non-attribution. Every synthetic utterance is an artificially generated sample. It is not something the corresponding speaker said, and must not be presented as their view, intention or endorsement. Identifying metadata that the research does not need, including speaker names and source-speaker mappings, has been removed.

Takedown. Rights holders and individuals whose voice appears here, or their authorised representatives, can request removal; valid requests are dropped from subsequent distributions.

Citation

@inproceedings{hear2026,
  title     = {HEAR Who Said What: Unlocking Speaker-Attributed Reasoning
               via Counterfactual Voice Grounding},
  booktitle = {Proceedings of EMNLP},
  year      = {2026}
}

License

Released under CC BY-NC 4.0. The underlying corpora keep their own licenses, AMI and ICSI are distributed by their respective consortia, and VoxMM is built from license-free YouTube material. Please honour those terms for any redistribution.

Downloads last month: 64

Total file size:

142 GB

Request access to HEAR and CASH-60K