Speech-DRAME-EvalBench — Metadata, Mel Features, and Reconstructed Audio

This release contains the 23,686 annotated experimental records used in the paper Speech-DRAME: Human-Aligned Evaluation for Speech Role-Play. It includes 8,272 archetype and 15,414 realism records. Additional internal development data, the 1,250 unannotated future generation prompts, model checkpoints, and original source waveforms are excluded.

This variant includes metadata, compressed mel features, and pre-rendered 24 kHz BigVGAN-v2 FLAC reconstructions. The FLAC files are vocoder reconstructions, not original recordings.

The companion Anuttacon dataset contains the same record IDs, split assignments, annotations, and mel features, without pre-rendered audio.

Splits

Perspective	Split	Records	Meaning
Archetype	train	6,773	Evaluator training
Archetype	test	1,499	Held-out role × scene combinations
Realism	train	10,990	Base evaluator training
Realism	test_base	2,000	Benchmark-distribution test; 210 positives, 832 role and 958 context negatives
Realism	realrec_adaptation	1,080	Additional real-recording training data
Realism	realrec_clean_test	1,084	Clean held-out human acting; speakers disjoint from adaptation
Realism	benchmark	260	Separate delivery-mismatch benchmark; 13 TTS systems × 20 Mandarin scenes

The real-recording pool is partitioned, not duplicated: 1,080 + 1,084 = 2,164. The final adapted realism evaluator uses 10,990 + 1,080 = 12,070 training records. The real-recording and delivery-benchmark splits are Mandarin-only. Archetype and the base realism splits contain English and Mandarin. Base realism is mixed human/TTS audio. Clip disjointness does not imply character disjointness.

Download and load

Metadata are JSON Lines. Asset files are bundled in uncompressed tar shards of approximately 256 MiB; each asset entry names both its archive and extraction path. Download the repository and extract only the desired assets:

from huggingface_hub import snapshot_download
from pathlib import Path
import json
import tarfile
import numpy as np

root = Path(snapshot_download("ftshijt/speech-drame-evalbench", repo_type="dataset"))
# Use Python 3.12+ for the safe tar extraction filter.
extracted = Path("speech-drame-extracted")
extracted.mkdir(exist_ok=True)
for shard in sorted((root / "assets").rglob("*.tar")):
    with tarfile.open(shard) as archive:
        archive.extractall(extracted, filter="data")

records = [json.loads(line) for line in
           (root / "metadata/realism/realrec_clean_test.jsonl").read_text().splitlines()]
mel = np.load(extracted / records[0]["audio"]["spectrogram"]["path"],
              allow_pickle=False)["mel"]

For direct playback after extraction:

import soundfile as sf
waveform, sample_rate = sf.read(extracted / records[0]["audio"]["synthesized"]["path"])

Each tar contains regular files only. manifests/assets.jsonl records the SHA-256 and archive membership of every asset; manifests/files.sha256 covers all release files except itself. Metadata can also be loaded using the archetype or realism configuration with the Hugging Face datasets library.

Annotation and provenance notes

Archetype annotations include Audio Quality, Human-likeness, Appropriateness, and Content Pass. Realism retains the eleven annotation sub-dimensions; the paper's evaluator predicts ten, excluding Semantic Match as a separate output. Per-rater arrays and missing/gated values are preserved; missing ratings remain JSON null. Use masking for the paper's primary correlations. The base realism training set contains 31 two-rater records; evaluation records have at least three raters.

Positive/negative construction is separate from audio provenance. Role/context negatives change text and retain their anchor clip's audio, which may be human or TTS; delivery negatives resynthesize audio. In test_base, each served record has a distinct source clip, with no second matching positive record in that set.

The old is_real_audio field was unreliable and is preserved only as legacy_is_real_audio, not as ground-truth provenance. source_type is human for the real-recording pool and tts for archetype/delivery-benchmark records; it is unverified for the mixed base realism records pending a corrected clip-level provenance mapping. Existing nonempty source attributes are retained; missing source information is not fabricated. Every distributed FLAC, where present, is a vocoder reconstruction regardless of source type.

Reproduction and limitations

The exact six canonical experimental metadata inputs were checked against the paper inventory. Record IDs, annotations, and scientific context are retained; split names, asset references, and provenance warnings have been normalized for public use. Public metadata input hashes and counts are in release.json. Mel values, dimensions, FLAC headers, and all derived asset checksums were checked before publication. These files reproduce release-native masked evaluation; legacy analyses that also floor model predictions require additional saved prediction artifacts. Code and model weights are separate releases.

vocoder/ contains the original mel frontend settings (100 bands, 24 kHz, 1024-point FFT, 256-sample hop) and the BigVGAN-v2 model identifier. Mel features and reconstructions retain voice-identity information and do not eliminate re-identification or voice-cloning risk. Scores reflect the bilingual annotator populations and can encode stereotypes. The paper reports strong benchmark alignment but poor synthetic-to-authentic transfer; this evaluator is not a validated general-purpose judge of human acting.

License and citation

Author-controlled contributions are released under CC BY-NC 4.0. Underlying source rights, attribution, and third-party model/service terms remain with their owners; this release does not relicense third-party works. Feature conversion does not itself establish redistribution rights. See LICENSE.md and SOURCE_NOTES.md. No additional research-only restriction is imposed.

• Paper
• Project
• Evaluation code

@article{shi2026speechdrame,
  title={Speech-DRAME: Human-Aligned Evaluation for Speech Role-Play},
  author={Shi, Jiatong and Han, Jionghao and Lu, Yichen and Pascual, Santiago and
          Wu, Pengfei and Cui, Chenye and Watanabe, Shinji and Weng, Chao and Zhou, Cong},
  journal={Transactions on Machine Learning Research},
  year={2026},
  url={https://openreview.net/forum?id=m9IP9ECrk2}
}