Beam-2 with a temperature-fallback ladder made the final-model warmup
run all three temperature retries under beam search before the server
bound its port, so /health was refused for ~8 min and STT was down that
whole time on every roll (and hinted at slow per-utterance decodes).
Production now runs the accurate large-v3-turbo model greedily at a
single temperature, keeping the proper-noun priming prompt that fixes
names like Amy/Córdoba - fast startup, fast decodes, accuracy intact.
Beam stays env-tunable (HERMES_STT_FINAL_BEAM_SIZE) for a future pass;
serve-before-warmup is a recommended follow-up so cold start never
blocks readiness.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BvMSXH8VH2tMWXanb8SJdf
Proper nouns (Córdoba, Cancún) and dropped words came from decoding the
committed transcript with the small model. The image already ships
large-v3-turbo, so the final decode now uses it with beam_size=5, a
temperature fallback ladder, and a proper-noun/accents initial_prompt
that fixes first-pass capitalization and diacritics across EN/ES/RU;
the rolling previews stay on tiny at greedy so the on-the-fly feel is
unchanged. The accurate decode runs in the speculative predecode during
the end-of-speech silence and is cache-reused at commit, so perceived
latency stays low. All decode knobs are env-overridable for on-device
tuning (beam/temperature/prompt), with small as the guaranteed-present
rollback if turbo underperforms on the Jetson. 206 STT tests pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BvMSXH8VH2tMWXanb8SJdf
