L'électron rare
966ff731bd
Server-side fuzzy alias table applied to the final STT transcript before forwarding to /voice/intent. Closes the metier-vocabulary gap of generic STT models (U-SON, LA 440 Hz, Zacus, ...) without needing per-request prompt biasing — which moshi-server's Rust runtime does not expose anyway. Loader is soft on pyyaml (degrades to no-op if missing) and on the yaml file itself (logs + skips on parse error). Patterns are case-insensitive with auto word-boundaries, with a (?...) escape hatch for advanced regex. Per-puzzle entries layer on top of globals — firmware passes puzzle_id in the hello frame; legacy clients omit it and get just the global polish. Loaded once at module import; restart voice-bridge to reload. Unit tested on 5 representative transcripts (eu son nez -> U-SON, diapazon -> diapason, le rasus -> Zacus, le la 440 hertz -> LA 440 Hz).
63 lines
2.1 KiB
YAML
63 lines
2.1 KiB
YAML
# STT post-correction fuzzy aliases for Kyutai (and any) STT output.
|
|
#
|
|
# The Kyutai 1B en_fr model does not know Zacus metier vocabulary
|
|
# (U-SON, the 440 Hz tuning fork context, custom NPC names, etc.) and
|
|
# the moshi-server Rust runtime does not expose any per-request
|
|
# TextPrompt mechanism we could use to bias the LM (cf.
|
|
# docs/voice-pipeline-roadmap.md). So we rewrite the transcript on the
|
|
# server side before forwarding to /voice/intent.
|
|
#
|
|
# Loaded by tools/macstudio/voice-bridge/main.py at boot via the
|
|
# `STT_ALIASES_PATH` env var (default: this file). Hot reload by
|
|
# restarting voice-bridge.
|
|
#
|
|
# Schema:
|
|
# global: # applied to every transcript
|
|
# - from: "regex pattern" # case-insensitive, word boundaries auto
|
|
# to: "replacement"
|
|
# per_puzzle: # keyed by puzzle_id (firmware passes it in hello)
|
|
# P1_SON:
|
|
# - from: "..."
|
|
# to: "..."
|
|
#
|
|
# Patterns are matched case-insensitively. Word boundaries are added
|
|
# around the pattern unless it starts with `(?` (escape hatch for
|
|
# bespoke regex). Replacements preserve case of the first matched
|
|
# character.
|
|
|
|
global:
|
|
# Professor name — the model often hears it as "the rasus" / "le rasus".
|
|
- from: "(?:le |la )?rasus"
|
|
to: "Zacus"
|
|
- from: "professeur ?(?:le |la )?rasus"
|
|
to: "Professeur Zacus"
|
|
|
|
# Common French STT polish — these are not Kyutai-specific but help
|
|
# everyone (e.g. whisper.cpp fallback would benefit too).
|
|
- from: "vue son"
|
|
to: "U-SON"
|
|
- from: "(?:eu|u) son nez"
|
|
to: "U-SON"
|
|
- from: "(?:eu|u) son(?:'s)?"
|
|
to: "U-SON"
|
|
- from: "husson(?: n'est)?"
|
|
to: "U-SON"
|
|
|
|
per_puzzle:
|
|
# Puzzle U-SON — frequency / tuning fork
|
|
SCENE_U_SON_PROTO:
|
|
- from: "(?:le |la )?la (?:quatre|4) ?cent ?quarante"
|
|
to: "LA 440"
|
|
- from: "(?:le |la )?la (?:quatre|4)40 ?(?:hertz|hz)"
|
|
to: "LA 440 Hz"
|
|
- from: "diapazon"
|
|
to: "diapason"
|
|
- from: "fréquence fondamentale"
|
|
to: "fréquence fondamentale"
|
|
|
|
# Reserve slots for the other scenes — populate as playtest reveals
|
|
# actual STT failure modes per puzzle.
|
|
SCENE_LA_DETECTOR: []
|
|
SCENE_MORSE: []
|
|
SCENE_COFFRE: []
|