- 已装运
- 2026年9月25日 01:19 UTC
- 作者
- Kamo
- 提交
- 5a8dc76
The probe runs inside the pod and prints one JSON document for register.py: a TTS round trip per voice and language (recall >= 0.8), the batch STT row, a realtime transcription session, and with --concurrency N that many sessions beside a ~60 s batch (p95 end-of-speech to transcript <= 4 s). Against a local Speaches 0.9.0-rc.3 (same lock, pinned models) three things differ from the plan and are fixed here: - the phrases drop the brand name: Whisper hears "Kamo" as Commo, Carmo, Camus or TAML, which failed Spanish on every run (recall 0.77) and left Piper one word from failing; brand-free phrases scored 1.00 in 16/16 runs; - the realtime session streams 100 ms frames at real-time pace: a burst made the server close the same turn ten times (duplicate commits, "item already exists" error events, the turn transcribed again for each); - the session joins every completed turn, keeps server error events for the detail instead of failing on the first, and commits the buffer itself only when all audio is sent and nothing is being transcribed. p95 is a true nearest rank (ceil), and the load block reports the batch detail.