Map the interaction stages
A recording may include automated menus, queue announcements, hold music, ringing, agent conversation, and post-call silence. State which stages belong in the dataset and why. For a dialogue model, long queue regions may be irrelevant; for interaction segmentation, those same boundaries can be part of the target behavior.
Define region labels and boundaries
Write rules for speech activity, automated prompts, music, tones, and silence, including mixtures. Distinguish an intentionally quiet interval from missing or failed capture where the source supports that decision. Do not label every low-energy segment as silence without checking whether a quiet participant is speaking.
Keep timing and counts interpretable
If non-conversational regions are trimmed, preserve the offsets or segment map needed to connect annotations to the source. Report retained conversation duration separately from original recording length. State whether repeated menu prompts are counted as independent speech examples or grouped as recurring source material.
Inspect representative transitions
Listen around the start of agent speech, resumed calls after hold, and the final disconnect. These transitions can expose missing sides or incorrect trimming. Share your retention rules and task priorities with Fiund so a prospective source and preparation plan can be reviewed against the required interaction coverage and project handling scope. Choose a call with two hold periods and verify that the conversation resumes at the expected point in every derived artifact. Check whether an automated announcement has been mislabeled as the agent. This small review exercise can expose segmentation mistakes that distort both usable-duration counts and the apparent sequence of conversational turns.
Review checklist
- Define included interaction stages
- Specify speech, prompt, music, and silence labels
- Preserve trim offsets and duration counts
- Inspect hold and connection boundaries