luvox-ai/HalluAudio-SlotQA-v2

LuVox HalluAudio SlotQA v2

Each audio has six multiple-choice rows: action, device type, device/scene name, location, target value, and attribute. Slot 1 uses A–I; slots 2–6 use A–G. Options are shuffled deterministically per audio and slot.

acceptance_rejection is assigned per slot row. acceptance means the label supports a clearly extracted value for that slot. rejection means the slot is not mentioned, the speech is unclear, or there is no human speech. Configs follow source_type-level[-variant...], such as synth-L1, record-L4, or record-L1-multi_devices. The original collection remains in source_config_name. region uses north, central, or south; it is null when the available annotation does not establish a region. Bắc Trung Bộ is grouped as central. This export covers speech-present labeled L1, L4, and L6 audio; invalid_noise remains excluded.

Step prompt variant — record-L6-step

This config contains 84 paired prompt variants of record-L6 (14 recordings × six slots). Audio bytes, transcript, all six option groups and their order, correct choice keys and other annotations are identical to the source. Only prompt and config_name change. Original configs remain available. The dataset now has 414 rows and 45 unique recordings; tables below count both prompt variants as rows, while audio counts deduplicate recordings. Do not treat paired variants as independent recordings.

The header’s instruction to output only one letter is replaced by: “Chỉ trả lời cho phần được hỏi ở cuối prompt, dựa trên nội dung nghe được.” After the six option groups, the requested slot is asked explicitly, followed by:

Đáp án của [1] là gì?

Hãy suy nghĩ từng bước. Dòng cuối cùng chỉ ghi một đáp án: A, B, C, D, E, F, G, H hoặc I.

Slots [2]–[6] instead use: “Hãy suy nghĩ từng bước. Dòng cuối cùng chỉ ghi một đáp án: A, B, C, D, E, F hoặc G.” The instruction changes response format; evidence and rejection-option meanings are unchanged.

Scoring this variant: extract the last nonempty response line, strip surrounding whitespace, and require exactly one uppercase choice key valid for the requested slot. Compare that key with expected_result using normalized match. Earlier explanation is not scored; do not compare the complete response directly with the key or extract a letter from the explanation. Existing scoring values stay unchanged; the variant-specific extraction rule is documented here and in dataset_summary.json.

from datasets import load_dataset
ds = load_dataset("luvox-ai/HalluAudio-SlotQA-v2", "record-L6-step", split="test")

WHAM people — record-L6-wham-people-step

Ten WHAM recordings × six slots = 60 rejection examples. These use WHAM ambience directly, with no added TTS, question speech, reverb or SNR mixing. Target: speech-like background conversations whose semantic content is insufficient for reliable slot extraction. All six expected choices mean “không xác định được vì không nghe rõ nội dung người nói”; they are distinct from no-speech and clearly-heard-but-not-mentioned choices. Candidate people/L6 labels are based on machine-generated WHAM annotation hints, not independently verified by a manual listening audit of all recordings. Treat these labels as annotation-derived targets, not verified transcripts. Background may also contain dishes or other incidental sounds; the recordings are not established as Vietnamese speech. Transcript stays empty, gender unknown and region null.

The prompt matches record-L6-step: all six option groups are present, only the requested slot changes, and the final line must be one choice key (A–I for slot [1], A–G for [2]–[6]). Use the last-line scoring rule above. Options are deterministically shuffled separately for each recording and slot using SHA256 of audio_id and slot identity; expected keys are recomputed after shuffling. Repeated letters across slots are allowed. All six rows share identical audio bytes. Keep them together when splitting data.

Audio is mono PCM16 WAV at 16 kHz, under 30 seconds. Whole excerpts are retained, normalized toward −20 LUFS (FFmpeg EBU R128, dual_mono=false); peak limiting and/or gain reduction is used if needed for ≤−1.2 dBTP headroom. No inferred transcript is created. Provenance and measurements identify source cache fingerprint, noise recording, source/output hashes, annotation hints and loudness processing. Validation records counts, keys, formats and shuffle checks.

WHAM attribution: Wichern et al., WHAM!: Extending Speech Separation to Noisy Environments, MERL and Whisper; original dataset and CC BY-NC 4.0 terms. Cached excerpts: nguyenvulebinh/wham; automatic hints: luvox-ai/audio_noise_nvlb_wham_annotated. Included WHAM noise remains subject to its noncommercial attribution terms. This config changes loudness and adds SlotQA prompts; other configs retain their source terms.

WHAM people, plain prompt — record-L6-wham-people

60 examples, sharing exactly the same ten recordings as record-L6-wham-people-step. Audio bytes, option order, answer keys, transcripts and all other annotations are unchanged. Only prompt and config_name differ. No additional recordings are introduced.

The prompt follows original record-L6: all six option groups remain present, the header says “Với phần được hỏi ở cuối prompt, chỉ trả lời một chữ cái tương ứng với lựa chọn bên dưới.”, and the final question is “Đáp án của [n] là:”. There is no step-by-step or final-line instruction. Score the entire response as a single choice key using normalized match; explanation is not requested. A–I apply to slot [1], A–G to slots [2]–[6].

These paired plain/step configs use the same deterministic shuffle, allowing comparison of prompt instructions while preserving audio and options. Keep identical audio_id/slot_id pairs together in data splits. The WHAM annotation-derived L6 labels, lack of independent listening verification, source provenance, processing and licensing terms described above apply equally to this config. Plain-variant validation.

Dataset counts

By level

Level	Audio	Rows	Acceptance	Rejection
L1	20	120	50	70
L4	1	6	0	6
L6	24	288	0	288
Total	45	414	50	364

By slot

Slot	Acceptance	Rejection	Total
action	15	54	69
device_type	10	59	69
device_or_scene_name	13	56	69
location	2	67	69
target_value	5	64	69
attribute	5	64	69

By config

Config	Audio	Rows
record-L4	1	6
record-L6	14	84
record-L6-step	14	84
record-L6-wham-people	10	60
record-L6-wham-people-step	10	60
synth-L1	20	120

By region

Region	Audio	Rows
central	5	30
north	15	90
unknown	25	294