Describe the conversation you need
Separate customer support, outbound sales, appointment handling, and other scenarios. State whether conversations must be genuine historical calls or whether commissioned role-play is acceptable. Include desired call lengths and exclusions such as automated messages, hold music, or unanswered calls.
Verify the original channels
Ask what the original recording contains and how exported tracks are produced. Two channels may duplicate the same mixture rather than separate participants. If your model needs isolated speakers, review that property directly on representative examples instead of relying on a filename or format label.
Define transcript conventions
Specify whether text should preserve disfluencies, repetitions, interrupted words, and overlap. Decide on speaker attribution and timestamp granularity. If multiple languages appear within a call, describe how they should be labeled and how language-level volume should be counted.
Review privacy and permitted use
Describe required removal of personal details from both audio and text. Ask what source permissions support the proposed project and which restrictions remain. Keep sample review separate from approval of the full collection and its commercial terms.