Start from the operational task
List the domains and call reasons that matter, such as account support, product troubleshooting, delivery changes, or appointment scheduling. Avoid treating every customer-service recording as interchangeable. A corpus dominated by order status calls may contain little evidence for complex technical troubleshooting even if both are labeled support.
Define vocabulary evidence
Request representative terms, abbreviations, product names, and spelled sequences with supporting audio segments. Keep vocabulary labels tied to what is actually audible rather than assumed from a transcript. State whether uncommon terms require specialized review and how uncertainty should be recorded when the recording is unclear.
Sample context around terminology
A short isolated keyword does not capture how callers explain problems or how agents ask clarifying questions. Preserve enough surrounding speech for the intended task. Include paraphrases, mispronunciations, corrections, and terms used in different meanings when these occur in the source and remain within the reviewed project scope.
Measure coverage without overclaiming
Count calls and independent speakers containing important term families, not just repeated word occurrences. Review a pilot to distinguish a naturally occurring vocabulary mix from scripted repetition. Send Fiund a prioritized domain lexicon, example intents, and coverage requirements so sourcing feasibility can be discussed against concrete evidence. One useful pilot case contains a product code that sounds like an ordinary word. Review whether the reference is grounded in audible spelling or inferred from context. Keep both the uncertainty and any supporting source terminology visible, since a confident but unsupported correction can turn a transcription reference into an unreliable target.
Review checklist
- List target domains and call reasons
- Attach audio evidence to specialist terms
- Retain useful conversational context
- Count independent speakers and calls per term family