Home/Quality
Quality assurance
Quality before quantity.
Speech data is only useful if it’s right. Every project runs against written acceptance criteria, and nothing is delivered until it meets them.
Checks
What we check.
The checks for each project are agreed at scoping. Depending on the project, they can include:
- Audio quality clipping, distortion, and recording level
- Silence empty or truncated recordings
- Background noise against the conditions the project calls for
- Duplicate detection repeated or reused recordings
- Language validation the recording is in the required language
- Dialect validation the speaker’s dialect matches the project
- Transcript validation accuracy against your transcription guidelines
- Metadata validation speaker and session fields are complete and consistent
- Human review a separate reviewer approves or returns every item
How review works
Two people touch every item.
Contributor
A verified speaker records, or a dialect-matched annotator transcribes and labels.
Reviewer
A separate reviewer checks the work against the guidelines, then approves it or sends it back with notes.
Rework, not rejection
Returned items are fixed and re-reviewed. Only approved data goes into your delivery.
Judge the quality yourself.
A pilot gives you a representative sample to check against your own model and criteria.