Describe the usable unit
Audio, speaker turns and any related resolution records. Explain whether the archive preserves the connections between records and the process they represent. File count alone does not show coverage, quality or usefulness.
Ask this before preparing a sample
Ask which languages, recording quality and interaction types are in scope, and who bears transcription and review costs.
Review sensitive and restricted elements
Recording permissions, voices, background speech and personal information. Check the actual contents, including hidden fields, metadata, embedded files and related records. Data minimization and de-identification are separate from confirming the right to license the material.
Evidence to bring to a fit discussion
Prepare a schema or category-level description, date range, approximate volume, export format, known gaps and exclusion list. Clearly distinguish measurements from estimates. Do not upload the underlying archive simply to obtain a preliminary opinion.
Compare the proposed use
Ask whether the provider wants training examples, evaluation material, retrieval content or a simulated environment. These are different purposes and should not be bundled into vague unrestricted language. Confirm downstream recipients, duration and any right to sublicense.
Tests to agree before a sample
- Check that a recording is complete and that its associated outcome is correctly linked.
- Identify whether speaker labels are inferred or verified.
- Document noise, language, channel and transcription limitations without claiming perfect accuracy.