Access to anonymized OpenBibleTTS listening-study ratings and original stimuli
Dear OpenBibleTTS authors,
We are building a multilingual meta-evaluation set to validate an audio-language model as a judge of same-text TTS pairs. We are interested in your human listening study for Hausa, Swahili and Turkish.
Your paper describes three native-speaker raters evaluating five TTS systems on 60 texts per language. The model cards link to davidguzmanr/open-bible-surveys, but the repository currently returns 404. Could you point us to the public annotation export or advise how to request access?
We need the exact synthetic audio clips used in the listening study, original transcripts and stimulus IDs, system and speaker/reference identifiers, and per-clip scores with anonymous rater IDs. An anonymized HumanSignal export with a stable audio mapping would be sufficient; we do not need any personal contact details of annotators.
We aim to select at least 50 distinct same-text synthetic-vs-synthetic pairs per language, then freeze five pairs for an initial pilot. We would derive preferences from the actual individual scores, exclude natural references, and preserve the original evaluated audio rather than re-synthesizing it.
Please also clarify the terms for local internal evaluation and for publishing a derived pair manifest, separately from redistributing audio or ratings. We will not assume the training-corpus license covers all evaluation artifacts.
Thank you.
Hi @hailinyu ,
Thank you for your interest in OpenBibleTTS and for explaining your intended use.
The davidguzmanr/open-bible-surveys repository is currently private, which explains the 404 error. We apologize for the confusion caused by linking to it in the model cards.
Unfortunately, we cannot share the listening-study ratings or original stimuli at this time, including for internal evaluation. We are reserving these artifacts for a planned multilingual MOS dataset, with additional languages and evaluation data, and potentially for a shared task in the next edition of the VoiceMOS challenge.
We plan to release the expanded dataset publicly in the coming months, although we cannot confirm a release date yet. As part of that release, we intend to clarify the terms for using the evaluation artifacts.
We hope it will be useful for your work, and thank you for your understanding!