- Shipped
- October 26, 2026 at 8:05 PM UTC
- Author
- Kamo
- Commit
- dc2c0ca
- E5 {"stream": true}: the audio is handed on chunk by chunk as the provider sends it (ElevenLabs **************** the candidate walk, so an empty answer still fails over before anything is sent; metered by characters as before. A family that cannot stream keeps the whole answer. Without the field nothing changes. Calls waited 950 ms (p50) for a whole sentence before its first audio. - TtsQualityRanking (D1, owner 2026-10-26: best quality among the org's configured and enabled providers, chosen automatically): with no TTS order the best VERIFIED model by published conversational quality and realtime fitness comes first, provider-neutral; unknown models tie with Kamo's in-cluster voices, so the in-cluster default still decides among them. An org's own order and an AI member's chosen voice still come first.