Skip to main content

Record a speech corpus

  1. Projects → Create project. Name it on the Project Name tab.
  2. On Labeling Setup, open the Voice Collection group and pick a template for your elicitation style:
    • Read aloud (voice corpus) — scripted prompts, read verbatim
    • Prompted question (open-ended speech) — spontaneous answers
    • Spontaneous speech (long-form monologue), Listen and repeat, Translation to speech, and more
  3. On Data Import, upload one task per prompt — a CSV whose column name matches the template's $variable: text for Read aloud, question for Prompted question, topic for Spontaneous speech. Save.
  4. Audio QC is on by default for templates that record with <AudioRecord> (all of the above): the tag auto-enables the Settings → Audio QC tab and its Run automated quality checks toggle. (Read-aloud with speech-text alignment lives under Audio/Speech Processing — it aligns existing audio with <AudioTextAlign> and doesn't record, so it has no Audio QC.) Tune the Pipeline, Recording environment, and Corpus purpose & validation cards there — the audio profile sets the same bar for recording hints, automated checks, and validation.
  5. Speaker profiles are enforced automatically on audio-QC projects: before their first recording, each contributor must complete the speaker-profile step (every field may be left blank and consent declined — the deliberate choice is what's required). The server re-checks on every submission.

Go deeper​