It starts with a specification
Before anyone records, the dataset is designed: language, speaker counts and quotas, speaking style, environment, audio format, transcription depth, metadata schema and licensing scope. Ambiguity here becomes rework later.
Recruitment and screening
Speakers are recruited against the written profile and screened for language, accent, voice suitability and recording conditions. Consent is captured at this point, not afterwards.
Recording, then a pilot
A pilot batch is recorded and reviewed so conventions, formats and quality thresholds are agreed while changes are still cheap. Full collection then runs with monitoring and progress reporting.
Transcription, annotation, QA, delivery
- ▸Transcription to agreed conventions
- ▸Segmentation, speaker attribution and labels
- ▸Audio, transcript, speaker and metadata checks
- ▸File-format and completeness validation
- ▸Packaged delivery with manifests and documentation