EarningsCallVoice · listening study
Which answer sounds like the reference speaker?
Listen to the reference, read the analyst's question, then compare nine anonymous answers. Rank them from most likely to least likely to have been spoken by the person in the reference.
Instructions and what to expect
Each set contains one real answer and eight voice clones. Judge how likely it is that the reference person spoke each answer. The answer transcript is hidden. Do not assign separate ratings for word accuracy, delivery quality or naturalness.
Listen to every answer before ranking. Use 1 for the most likely, then 2, 3, and so on. Give tied answers the same rank, with no gaps: for example, 1, 2, 2, 3. You can replay any clip.
Use headphones, a quiet room, a comfortable volume, and normal playback speed. Five sets make a session. The original reviewer spent around 20 hours on all 100 sets, so take breaks and contribute as many complete sets as you wish.
Please judge from the audio without looking up model identities or previous rankings. If a clip cannot play, leave that set incomplete and report its set and candidate letter.
Your work stays in this browser until you click Submit labels. Browser storage can be cleared, so download a backup regularly. Exporting a CSV or backup does not submit anything.
A code is created for you. Keep it when returning to the study; do not enter your name or email. After entering another code, click Open participant. To move to another browser or device, export and import a backup.
Analyst's question
Reference speaker
Return to this reference whenever you need to compare the voices.
Save and share your work
You can contribute after any complete set. Click Submit labels to send your participant code, completed rankings, confidence scores and notes to the researchers for this study. Results are stored privately in the study owner's Hugging Face account. No email address or login is needed.
Submission needs an internet connection. You can keep labeling and submit again later; repeat submissions are treated as updates. A CSV keeps completed sets, while a backup also keeps unfinished work.