Raw hours is a weak metric for speech corpora. Coverage across speakers, settings and speaking styles is what moves word error rate in production.
Prosody shapes meaning
Japanese is mora-timed with pitch accent. Homophones are distinguished by pitch contour and context. Models trained on read speech often underperform on spontaneous speech, where reductions, fillers and sentence-final particles dominate.
Dialect is not an edge case
Regional varieties differ in vocabulary, accent and grammar. A corpus that spans only standard speech leaves a large share of real users underserved. Dialect labels should be recorded at collection time, not inferred later.
Consent and voice rights
- Recording consent must cover AI training and, for TTS, voice synthesis explicitly.
- Call-center audio requires consent from both parties or fully scripted participants.
- Minors and elderly participants require additional safeguards.
Recording conditions
Studio audio suits synthesis. Recognition needs noise, distance and device diversity. Specify sampling rate, channels and environment tags, and sample them separately in QA.
General information, not legal advice.