Heiheihaha17/DiSH-Bench-static-v2

DiSH-Bench static v2 — test release

A simulated multi-receiver spatial speech benchmark, with aligned single-receiver and multi-receiver questions over the same physical scenes. This release contains 122,256 test questions, 2,000 physical scenes, 4,400 original four-channel FOA recordings, and 200 rooms. There are 58,829 single-input questions, 53,518 two-input questions and 9,909 three-input questions. Questions share recordings; they are not independent acoustic scenes.

Source revision: unified_static_guarded_v2_20260930. Test split only. The source release audit passed. Packaging preserves all original test JSONL records and waveform bytes; asset roots are made portable. A Parquet index is added for convenient loading. See audit/package_validation.json and manifest.jsonl for the published-file checksums.

Tasks

Task	Questions	Task family
DOA	11,200	Directions of human speech and non-reference robot playback
RANGE	17,720	Human-speaker distance conditioned on an angular sector
TEXT_DOA	20,160	Direction associated with specified speech text
DOA_ASR	15,264	Transcription conditioned on direction
C1_SETLOC	7,600	Human sound-source localization set
C2_REGION_ASR	16,744	Human transcription within a world-coordinate region
C3_TEXTLOC	13,680	Localization associated with specified speech text
KW_ASR	19,888	Full human transcription selected by an utterance-prefix wake keyword

The stored prompt and output_schema define each question's exact output format. private/test/gold.jsonl provides answers for the original tasks; extensions/kw_asr_v1/private/test/gold.jsonl provides KW-ASR answers. Robot playback is excluded from human-target tasks. DOA instead includes non-reference robot playback; follow the question's explicit target policy.

KW-ASR extension

The kw_asr_prefix_v1 extension adds 19,888 KW_ASR questions (11,527 Edge / 8,361 Cloud), in 5,244 same-query receiver groups. All original 102,368 questions, prompts, scene/view/recording metadata, and 4,400 WAV files remain byte-identical. No new audio or physical scene is generated.

The provided keyword is a 1–3 word utterance prefix, taken from real recorded speech. For example, rest across matches rest across the fire to the; across the inside that utterance does not match. Exactly one distinct human speaker must begin an utterance with the prefix. Transcribe all their speech in the full supplied segment, including the keyword. Zero or multiple matching humans require "". Robot playback never qualifies. Matching ignores case/punctuation and uses whole words. A person's utterance may start anywhere in the mixture. Positive targets have at least two additional words.

Case counts: ambiguous=35, no_match=5,487, positive=8,879, robot_only=5,487. Report positive-target corpus WER separately from empty-response accuracy for each negative case; split by Edge/Cloud and receiver count. Multispeaker-prefix cases are naturally sparse (only 9 query groups), so report their counts and do not present them as a large representative sample. Robot-only negatives include single receivers that have removed their own playback; these views correctly have an empty answer but are not audible playback challenges. Playback text is known from the inherited public logs, so robot-only detection is a diagnostic rather than an independent ASR score.

extensions/kw_asr_v1/public/test/ stores the new inputs, prompts and pairs. extensions/kw_asr_v1/private/test/ stores answers and prefix-match provenance. The updated load_sample.py reads both the original and added questions. The default config contains all questions; base_static_v2 contains only the original tasks, and kw_asr contains only KW-ASR. See the reproducible builder and evaluation guide. The archived base manifest and upload proof are under audit/base_static_v2/; the root summary and manifest describe the combined release. New KW target-group IDs are opaque and do not expose human source IDs.

Load the index

from datasets import load_dataset

test = load_dataset("Heiheihaha17/DiSH-Bench-static-v2", split="test")
row = test[0]
print(row["sample_id"], row["task_id"], row["prompt"])

The Parquet index contains model-input metadata and prompts, with query_json encoded as a JSON string to support heterogeneous tasks. audio_paths are repository-relative strings, not automatically decoded Audio features. Download the snapshot to access audio and complete metadata:

from huggingface_hub import snapshot_download
import importlib.util
from pathlib import Path

root = Path(snapshot_download("Heiheihaha17/DiSH-Bench-static-v2", repo_type="dataset"))
spec = importlib.util.spec_from_file_location("dish_reader", root / "load_sample.py")
reader = importlib.util.module_from_spec(spec)
spec.loader.exec_module(reader)
item = reader.get_model_input(root, row["sample_id"])
# pip install soundfile; each array is [samples, 4], in the supplied slot order.
waveforms = [reader.read_audio(a)[0] for a in item["audio"]]
answer = reader.get_gold(root, row["sample_id"])

Use time_window_samples = [start, stop] as a half-open sample interval. Stored prompts display scene-relative times; cropped audio starts at start / 16000. Do not downmix four-channel inputs for benchmark evaluation.

Files and correspondence

• data/test-*.parquet: one row per question; no scoring answers.
• public/test/inputs.jsonl: canonical inputs, audio bindings, queries, windows.
• public/test/prompts.jsonl: exact rendered model prompts, keyed by sample_id.
• public/test/recordings.jsonl: recording IDs, receiver IDs, format and hashes.
• public/test/scenes.jsonl: scene IDs, robot poses, playback logs, room assets.
• public/test/views.jsonl: receiver subsets in each scene.
• public/test/pairs.jsonl: same-task, same-target correspondence across views.
• public/audio/test/: all 4,400 original WAV recordings, about 7.58 GB.
• public/scenes/: room meshes, materials and semantic definitions (600 assets).
• public/asset_roots.json: asset-root aliases resolved relative to snapshot root.
• private/test/: answers, source geometry, target views and diagnostic metadata.
• private/v1_to_v2_mapping.jsonl: test-only mapping from the parent release.
• provenance.json, summary.json, audit/: release provenance and checks.

This repository is public, including the directory named private: that name marks scoring metadata, not access control. Do not provide gold answers, human source geometry, or target IDs to the model. Use the stored prompt and its bound audio as the evaluation interface; full-precision metadata supports auditing.

scene_id identifies the physical scene; recording_id identifies a receiver's recording; view_id identifies a receiver subset; sample_id identifies a question. query_group_id joins comparable views, with target_group_id and reference_robot_id preserving target and coordinate-reference semantics. Do not assume a multi-input directional question changes its reference frame: follow the stored reference robot and prompt.

Official evaluation code

The repository now includes one scoring entry point for all eight tasks, using the published test metadata. The original seven tasks use the frozen original formulas and thresholds; KW-ASR retains its previously published scoring rules. Model-input prompts remain in their existing JSONL files and Parquet columns.

python -m pip install -r evaluation/requirements.txt
python evaluation/score.py --root /path/to/hf-snapshot \
  --predictions predictions.jsonl --output results --strict

Each prediction row contains sample_id and raw_text (the original model response). The default scores all 122,256 questions; --tasks legacy scores the original 102,368 questions, and --tasks KW_ASR scores the 19,888 new questions. Use --prediction-format native explicitly for existing structured unit-vector predictions instead of the degree pairs requested by the prompt.

Outputs include per-task metrics by Edge/Cloud and receiver count, per-target diagnostics, prediction reconciliation, and paired receiver comparisons. Missing predictions remain failures unless --strict rejects the incomplete submission. No cross-task total score or LLM judge is used. See the evaluation guide, scoring provenance, and full-split validation. All questions, answers, prompts, index rows, scene metadata and recordings are unchanged by this scoring-code addition. The preceding release proof and manifest are archived under audit/kw_asr_release_20261001/.

Acoustic protocol and limitations

All scenes are static and strictly synchronized at 16 kHz. FOA is ACN/SN3D in channel order [W,Y,Z,X]; receiver-frame X points front, Y left, Z up. World +Y is up. Robot yaw zero faces world -Z, and positive yaw turns toward world +X. Microphone and loudspeaker centers coincide. Each receiver's own playback is ideally removed; other robots' playback remains audible. R01 is silent and R02/R03 can play speech, as in the parent dataset.

Prompts round robot positions and query regions to one decimal place in meters, headings and directional cues to integer degrees, and timestamps to milliseconds. The inherited room-geometry/material description retains its source precision. Full-precision annotations are retained. Regional queries have a 0.2 m boundary guard; test RANGE queries have at least a 1 degree audited margin under actual and displayed poses. The underlying revision rebuilt 541 physical test scenes and 1,214 recordings; 3,186 recordings are retained from the parent release. Some region boxes extend beyond room walls; these remain geometric query boxes, and their recorded exceptions are included in provenance.json.

This release has no real multi-point recording validation, active receiver motion, environmental event task, clock drift, synchronization error, or segment loss. It is intended for controlled simulated evaluation. Generation provenance may name dry-speech, RIR or intermediate-mix paths: those intermediate assets are not included and are not needed to run the published test inputs. Source speech corpus IDs remain in the metadata; this dataset card does not assign a new license to upstream speech assets.