AnyEditBench

AnyEditBench is a benchmark for speech and singing voice editing under clean, simulated, and real-world acoustic conditions. It accompanies the NeurIPS 2026 paper AnyEdit: A Unified Framework for Speech and Singing Voice Editing with Real-World Environmental Consistency (project page).

Each item provides a source waveform, its original text (transcript or lyrics), and the target text after editing. A system is evaluated on how well its output matches edit_text while preserving speaker identity, prosody (melody for singing), and the background acoustics of the source recording.

The benchmark is distributed in two equivalent forms:

• Parquet (data/*.parquet): audio embedded, previewable in the dataset viewer and loadable with 🤗 Datasets.
• Zip (AnyEditBench.zip, ~110 MB): plain WAV files plus JSON metadata, no extra dependencies.

Composition

Condition (dataset_type)	Speech EN	Speech ZH	Singing EN	Singing ZH	Total
Clean	50	50	50	50	200
Simulated	50	50	50	50	200
Real-world	50	50	50	50	200
Total	150	150	150	150	600

• Edit types. The clean and simulated speech subsets each contain 30 single-span and 20 multi-span edits. All singing subsets and the real-world speech subsets contain single-span edits only.
• Simulated conditions. Clean sources are degraded with SNR ∈ {0, 10, 20, ∞} dB and RT60 ∈ {0, 0.5, 1.0, 1.5} s, excluding the fully clean and dry case (15 conditions). The condition of each clip is encoded in its filename, e.g. snr10_rt601.0; n stands for no noise or no reverberation.
• Real-world speech. 32 of the 100 clips carry background=clean and the rest background=noisy, as clip-level ambient tags.
• Audio. All 600 waveforms are mono, 16-bit PCM WAV at 16 kHz.

Sources

Subsets	Source	Upstream license
clean_speech_en, simulated_speech_en	LibriTTS	CC BY 4.0
clean_speech_zh, simulated_speech_zh	AISHELL-1	Apache 2.0
*_singing_*	SingVERSE	CC BY 4.0
real_speech_en, real_speech_zh	YODAS	CC BY 3.0

Note on singing clips. SingVERSE records the same sung phrase in several acoustic scenarios and pairs every noisy recording with one shared studio reference. Clean singing items whose filenames differ only in the scenario (e.g. Piano_Room vs. Shopping_Mall) therefore contain the same clean waveform. This is expected.

Subset names

Each subset is named {condition}_{modality}_{language}, e.g. clean_speech_en, simulated_singing_zh, real_speech_zh.

Fields

Field	Description
id	Stable id, {subset}_{index:03d}
audio	Source waveform (Parquet only; 🤗 Audio feature with embedded WAV bytes)
subset	Subset name (Parquet / JSON filename without extension)
language	EN or ZH
type	Speech or Singing
dataset_type	Clean, Simulated, or Real-world
edit_type	Single Span Edit or Multi Span Edit
original_text	Source transcript or lyrics
edit_text	Target text after the edit
audio_path	Path of the waveform inside AnyEditBench.zip, audio/<subset>/<file>.wav
background	clean or noisy, clip-level ambient tag
sample_rate	Sample rate in Hz (always 16000)
channels	Number of channels (always 1)

Usage

🤗 Datasets (Parquet)

from datasets import load_dataset

ds = load_dataset("Ririiiii/anyeditbench", split="test")                      # all 600 items
clean_en = load_dataset("Ririiiii/anyeditbench", "clean_speech_en", split="test")  # one subset

row = ds[0]
print(row["original_text"], "->", row["edit_text"])
wav, sr = row["audio"]["array"], row["audio"]["sampling_rate"]

To get the raw WAV bytes without decoding, use ds.cast_column("audio", Audio(decode=False)).

Zip (plain WAV + JSON)

wget https://huggingface.co/datasets/Ririiiii/anyeditbench/resolve/main/AnyEditBench.zip
unzip AnyEditBench.zip

AnyEditBench/
├── audio/<subset>/*.wav   # 600 waveforms
├── json/<subset>.json     # 12 files x 50 items
├── metadata.jsonl         # all 600 rows
├── README.md
└── LICENSE.md

import json
from pathlib import Path
import soundfile as sf

root = Path("AnyEditBench")
rows = [json.loads(line) for line in open(root / "metadata.jsonl", encoding="utf-8")]
row = rows[0]
wav, sr = sf.read(root / row["audio_path"])  # audio_path is relative to AnyEditBench/
print(row["original_text"], "->", row["edit_text"])

Texts are whitespace-normalized. For WER/CER, apply your usual text normalization and resample audio to the rate your ASR model expects.

Please cite the upstream datasets listed above when you use the corresponding subsets.

License

The benchmark annotations are released under CC BY 4.0. The audio remains subject to the licenses of the original sources; see LICENSE.md.