Consent, metadata, and QA built into the workflow.
Sonexis does not treat voice data as raw audio files. Every project is designed with consent, structure, review, and delivery readiness from the beginning.
Seven layers of quality and structure.
Consent workflow
Contributors agree to project-specific data usage terms before recording. Consent is task-specific, not a blanket platform agreement. Consent status is tracked and tied to each individual submission. A submission without confirmed, task-linked consent is excluded from project delivery.
Contributor screening
Contributors are screened by language, region, device, availability, and task fit before receiving assignments. Not every applicant receives tasks. Screening is designed around the specific requirements of each project.
Task design
Each task includes the scenario, language requirements, speaker structure, recording format, expected environment, and quality expectations. Ambiguity in task design leads to unusable submissions. We design tasks with precision.
Metadata structure
Every submission includes structured metadata that travels with the data from collection through delivery. Buyers receive data that is labelled, contextualised, and ready for integration.
QA review
Submissions are checked for audio quality, task accuracy, speaker validity, language match, consent status, duplication, completeness, and usability. Submissions that pass QA move into project delivery. Those that do not are rejected.
Rejection criteria
The following types of submissions are rejected and do not enter project delivery:
Delivery readiness
Datasets are delivered in structured formats based on buyer requirements: audio files, transcripts, metadata sheets, QA notes, manifests, and evaluation labels where required. The delivery format is agreed before collection begins.
Structured delivery, not a file dump.
Consent tracking
Consent-linked records attached to each submission, traceable per task and per contributor. Consent scope is defined per project.
Speaker and language metadata
Speaker profile, region, accent, and language tags for every recording.
Scenario context
The scenario, task intent, and expected behaviour documented for each recording.
QA-reviewed submissions
Only approved submissions enter the delivery batch. Rejections are excluded.
Known issue notes
Any flagged issues are noted in the metadata rather than silently included.
Agreed delivery structure
Audio files, manifests, metadata sheets, and transcripts in the format agreed before collection.
Raw unreviewed submissions, PII documents, or consent records identifying individual contributors are not included in standard delivery unless explicitly agreed under a separate legal and delivery scope. Only QA-approved, structured data enters the delivery batch.
Usage scope agreed per project
Each contributor agrees to task-specific data usage terms covering the agreed project scope, including AI training, evaluation, or benchmarking where applicable. Consent scope is defined per project and referenced in delivery metadata. Privacy review is available based on project scope. Data is structured for defined use, traceability, and delivery review.
Our contributor workflows are built for transparency: contributors know what they are recording, why it exists, and what usage terms apply. This traceability protects buyers and contributors alike.
Personally identifiable details that appear in recordings, such as names or contact references mentioned in conversation, are handled according to project-specific data management agreements. Known or detected PII indicators can be flagged in metadata where required, rather than silently treated as clean data.
Discuss QA requirements.
Tell us the QA standards your model requires, whether for ASR fine-tuning, voice agent evaluation, LLM training, conversational AI development, or benchmarking. We will scope the right review process and delivery structure.
Scope a Dataset