doteBench v1.0.2

doteBench 1.0.2 improves edit scoring and adds AuK with the official Prompt Enhancer.

• Missing candidate output remains in source-qualified scoring. Pitch and rate distinguish execution from success within the requested magnitude range.
• AuK uses cached Prompt Enhancer instructions for full-utterance generation.
• Updated reference tables and radar figures report the current evaluation results.
• Materialization validity follows data and audio contracts, independently of the software release label.

Install the attached wheel or source distribution. See the usage guide for setup and evaluation.

Audio assets are unchanged: use the audio bundle and its checksum. Common Voice 20.0 is obtained separately.

doteBench v1.0.1

doteBench 1.0.1 improves data materialization and compiler version tracking.

• Complete audio bundle with 1,027 content-addressed WAV files.
• Per-asset Common Voice conversion profiles reproduce all 226 reference WAVs.
• Materialization verification supports services launched through python -m.
• Compiler manifests use positive integer versions recorded in run metadata.

Validation: all 428 tests passed, and all Common Voice conversions and bundle WAVs matched their reference hashes.

Release assets:

• dotebench-1.0.1-py3-none-any.whl — Python wheel.
• dotebench-1.0.1.tar.gz — source distribution.
• dotebench-audio-bundle-v1_0.tar.gz — audio bundle.
• SHA256SUMS — artifact checksums.
• dotebench-audio-bundle-v1_0.tar.gz.sha256 — companion bundle checksum.

Follow Using doteBench to install, materialize data, generate audio, and evaluate results. See Data materialization for Common Voice and model setup, and the audio notice for attribution and bundle terms.

doteBench v1.0.0

doteBench v1.0.0

doteBench v1.0.0 releases a bilingual benchmark for precise speech editing: 2,081 English and Chinese cases across seven shards, with XML instructions, candidate adapters, deterministic data materialization, and a complete evaluation pipeline. The dataset and evaluation protocol identities are both v1_0.

Assets

• dotebench-1.0.0-py3-none-any.whl: installable Python package.
• dotebench-1.0.0.tar.gz: source distribution.
• dotebench-audio-bundle-v1_0.tar.gz: 1,026 content-addressed WAV files and their provenance and license notices.
• SHA256SUMS: checksums for the three assets above.

The audio bundle includes material from LibriTTS-R, FLEURS, Multilingual LibriSpeech, AISHELL-1, AISHELL-3, and speech generated with Qwen3-TTS VoiceDesign. The audio notice and provenance ledger record the applicable terms for each file. Additional Common Voice audio and generated assets are obtained during local materialization as described in the usage guide.

The data notice identifies transcript sources and their terms, including noncommercial conditions for WenetSpeech and Emo-Emilia portions. The third-party inventory covers adapted CC-BY-SA-3.0 speaker source, Praat GPL obligations for separately installed environments, and IndexTTS-2 model terms. Model checkpoints are obtained from their providers under the applicable terms.

Verification

The release source passed 398 automated tests, data validation across all 2,081 cases, compiler checks, provenance verification, wheel and source build audits, and an installed-wheel CLI smoke test. See README.md for installation and documentation links.