gmarti/EarningsCallVoice-Cloning-Listening

Voice Cloning Listening Study

Open the listening app or download the offline package.

This package contains 100 earnings-call listening sets. Each set provides a reference voice, an analyst question and nine anonymous candidate answers: one recorded answer and eight voice clones. All 1,000 WAV files are mono, 16 kHz, PCM-16. No answer key, previous rankings or participant responses are included. Please do not look up candidate identities while labeling.

Participate

Open the app, keep the participant code it assigns, and follow its listening instructions. Rank every candidate from most to least likely to have been spoken by the reference speaker. Ties are allowed; rank tiers start at 1 and have no gaps. Give a confidence score from 0 to 100. Complete as many sets as you wish; five sets make a session. The original reviewer spent roughly 20 hours on the whole package, including replays and labeling.

Progress saves in your browser. Export a JSON backup regularly and import it to move devices. Click Submit labels to send completed rankings, confidence scores, optional notes and your participant code privately to the researchers for this study. No email address or account is required. Submission needs an internet connection; CSV exports and backups stay on your device until you choose to share them. Avoid personal details in optional notes. You can keep labeling and submit again later: changed rankings are recorded as updates.

For offline use, unzip the download and open index.html. No installation, Git or Python is needed for listening. Python users can validate an export with python validate_results.py your_export.csv from the package directory.

Package and results format

Interface 1.0.1 adds private submissions. Study version 1.0.0, audio hashes, browser progress, backups and presentation order stay compatible. The package preserves the original audio and ranking question. It adds participant-specific item and candidate order, five-set navigation, partial exports, and separate browser progress for each participant. Orders remain stable when resuming with the same code and package version.

CSV records include the original anonymous candidate code and WAV checksum, the displayed candidate letter and set position, participant code, rank, confidence, optional notes, and package version. An export contains nine rows per complete set. Incomplete sets are kept in the JSON backup. Treat repeated exports by the same participant as versions of their labels, not new raters. The validator checks each file; merging or updating returns requires keeping the participant and set identifiers. CSV text that could be interpreted as a spreadsheet formula is prefixed with an apostrophe; JSON keeps the text as entered.

The public package changes presentation order and collection logistics; comparisons with the original review should record this package version. An overall rank does not separate speaker resemblance from other audible cues.

Source and attribution

Authentic clips and question text come from EarningsCallVoice: Core-100 v1.0.0, derived from FinCall-Surprise under Apache-2.0. SOURCE_NOTICE and SOURCE_LICENSE preserve its attribution and license. These files apply to the source material. The cloned answers were generated for the voice-cloning benchmark; model and source provenance are described in the benchmark article. Model names are intentionally omitted from the listening interface.

The browser interface and result validator are available under APP_LICENSE. This research package is for studying voice cloning and listener judgments. Please retain the source attribution and do not present cloned recordings as statements made by the original speakers.

Research by Gautier Marti and Hamdan Al Ahbabi, Khalifa University. The interface code is also in the Space repository. MANIFEST.sha256 lists the release file hashes; study.json contains only anonymous presentation data.