Define quality for the task
Clean audio is not always the only useful audio. A robustness evaluation may require difficult recordings, while reference annotation still needs enough intelligibility for credible labels. Specify which conditions are mandatory, acceptable, or excluded, and keep audio quality separate from whether the conversation content fits the domain.
Inspect several failure modes
Listen for clipping, dropouts, low level, heavy compression, background activity, echo, and missing speakers. Review each channel independently when present. A file that opens successfully can still contain long silent regions or an absent conversation side, so file validity should not substitute for listening checks.
Preserve source and processing details
Record the delivered format, sampling rate, channel count, and any resampling or enhancement. Converting a narrow-band source into a higher sampling rate does not recreate information absent from the original capture. Keep derived audio linked to its source so buyers can understand what changed during preparation.
Report distributions and exceptions
Summarize quality conditions across the sample and identify excluded or uncertain recordings. Avoid declaring an entire collection suitable from a few clean clips. Share the target conditions and listening-review requirements with Fiund so each potential source and processing plan can be assessed within the scope of the intended project. Listen to a segment from the beginning and another near the end of a long call using the same playback conditions. A source can change quality after a transfer or contain a late dropout. Record the location and nature of the issue so the review describes the recording rather than a single convenient excerpt.
Review checklist
- Define acceptable and excluded conditions
- Listen for missing sides and dropouts
- Record source format and transformations
- Report quality coverage and unresolved issues