Datasets:

Hayeonbang
/

VocalCoachBench

This directory publishes the final 26 October 2026 results for 13 models (9 open-weight, 4 closed API) and one no-audio majority reference. It accompanies VocalCoachBench on Hugging Face, the paper, accepted to NeurIPS 2026, and the project demo.

Dataset version: canonical515_physical_audio_1064refs. The released inventory has 515 recordings (A=207, B=308), 1,064 Top-3 expert references, 4,506 feedback blocks and 12,216 atomic claims (6,162 diagnosis, 4,593 correction, 1,461 strength). See the public release notes and the canonical Top-3 policy.

Files and provenance

leaderboard.json: unrounded ratios, definitions, coverage, model caveats and hash-only source provenance. Metric values, coverage and notes match the verified final export; internal source paths were removed for this release.
leaderboard.csv: byte-exact copy of the final CSV; metric columns are percentages. Empty cells mean not evaluated, while 0 means an evaluated zero.
provenance.json: source snapshot/receipt hashes, public file hashes, the publication transformation, and hashes of the supporting released annotations.
prompts/: eight byte-exact generation prompt copies with file and whitespace-stripped runtime hashes in prompts/manifest.json.

The internal receipt and experiment directories are not bundled. Their hashes record lineage; they do not make unavailable source files publicly reproducible.

Metrics and denominators

Metric	Definition and population
Top-3 macro-F1	Seven one-vs-rest classes over all 515 recordings and 1,064 references; average reference contributions within each recording. This is not F1@3 overlap.
Segment macro-F1	Seven-class macro-F1 on the fixed 262 evaluated segments, retaining invalid attempts.
Segment Match	Predicted class matches any accepted reference for each of the 262 segments.
Score ranking	Pairwise accuracy induced by quality scores; 189 triplets, 567 pair relations and 1,134 expert-reference attempts. Ties or missing scores fail.
Direct ranking	Accuracy of audio A/B decisions against both expert orders on the same 1,134 attempts; malformed outputs fail. This is a separate task from Score ranking.
Diagnosis strict hit	STRICT_HIT / 6,162 expert diagnosis claims.
Diagnosis hit	(STRICT_HIT + COARSE_HIT) / 6,162.
Prescription valid	VALID / 2,575 fixed slots (515 recordings × 5); missing prescriptions fail. This is not VALID / emitted prescriptions.
Diagnosis-to-prescription E2E	Valid prescription linked to a supported model diagnosis / 2,575 fixed slots. It measures supported advice, not demonstrated improvement in singing.

No overall composite rank is defined. The majority reference has only a Top-3 score. Human expert-vs-expert Top-3 analysis has 513 eligible recordings, since two recordings have one expert; the leaderboard's model denominator remains 515.

The 262-segment results use the fixed reference events released in annotations/segment_consensus_events.jsonl. Audio for 166 clips is packaged; the other 96 clips must be resolved from their source datasets. Use these exact events rather than selecting arbitrary rows or rebuilding a different consensus population. Frozen internal runtime manifests, model inference runners and judge runners are not bundled; the released events and available audio alone do not constitute a complete reproduction pipeline.

Evaluation and model policies

Primary structured results use strict_all_attempts: malformed, missing and backend-error outputs remain failures. Stored first responses are retained rather than replaced with successful retries. Top-3 references have equal total weight per recording. The seven public labels are PITCH, RHYTHM, DICTION, BREATH, PHONATION, TECHNIQUE and EXPRESSION; historical VOCALIZATION maps to PHONATION. Dataset-B annotation joins use physical-audio SHA-256, not filenames.

Claims results use the corrected515 format-reported extension: invalid Claims JSON emits no claims, and all fixed denominators remain. This differs from the older schema-penalized policy in historical protocol documents. The judge is openai/gpt-oss-120b; unresolved judge errors are zero in the verified cohorts.

Main uses structured JSON prompts for FunAudioChat, Kimi-Audio, MiMo-Audio, Qwen2.5-Omni, Qwen3-Omni, R1-AQA, GPT-Audio 1.5, Gemini 3 Flash Preview, Qwen3.5-Omni Flash and Qwen3.5-Omni Plus. Simple uses separate short Top-3, score, segment and overall-feedback prompts for Audio Flamingo 3, Music Flamingo and Covo-Audio, with their original text-extraction route. Simple overall feedback is decomposed into claims before judging; it is not the separate Natural task. The prompt manifest identifies the appropriate files; Main prompts must not be substituted for Simple.

Audio Flamingo 3 and Music Flamingo Direct use A audio, one second of silence, then B audio in one WAV with interval instructions. Their dynamic Direct wrappers are not included; the bundled Main Direct prompt expects separate labeled inputs. Covo-Audio Direct is N/A, and its 515 Claims outputs produced no usable claims. FunAudioChat Direct retains 18 failed first responses (549/567 parsed decisions). GPT-Audio 1.5 Direct uses the October 14 strict-first cohort; its other tasks reuse audited historical responses. Closed API snapshot/runtime equivalence and transport retries before each stored raw row are not established.

The prompt copies preserve frozen protocol text, including trailing whitespace. Their inclusion does not supply model adapters, system prompts, runtime wrappers, claim extraction/judge prompts, predictions or a complete reproducible inference and judging kit. Historical v5 examples on the demo are illustrations from a separate cohort; use this snapshot for the current aggregate results.