Gemma 4 cloud audio was infeasible (31b-cloud has no audio; E4B broken
upstream, no deploy host), so improve faster-whisper instead.
- Pin temperature=0.0 to disable the fallback ladder that re-decoded unclear
audio up to 6x (source of the 16-24s latency outliers); reject hallucinated
segments via avg_logprob/compression_ratio in the new pure _filter_segments.
- Adopt mikr/whisper-small-ro-cv11 (CT2 int8) via configurable voice.stt_model:
spike showed WER 24%->10%, numbers fixed at source, +0.33s p50 (in budget).
- Add tools/voice_stt_mine.py (log mining) + tools/voice_stt_spike.py (model
eval with diacritic scoring) + tests for the gate and miner.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Squashed branch: voice/dave-recv → master. Closes Pas 12 (DAVE E2E) and lands
voice-mode UX polish + verbal voice control on top of the Pas 1-10 scaffolding
already on master.
## DAVE E2E receive-side decrypt (e4f3177)
Vendored fork: discord-ext-voice-recv 0.5.3a+echo.dave1. Patches the receive
pipeline to handle Discord's mandatory DAVE encryption on voice gateway v=8.
- `_maybe_dave_decrypt`: uses davey.can_passthrough(user_id) as primary gate,
falls through to dave.decrypt for DAVE-epoch peers, drops on decrypt failure
without killing the reader thread.
- VAD fix: silero-vad v5+ requires exactly 512 samples; our 100ms window
(1600 samples) was silently raising ValueError → STT never fired. Now slice
into 512-sample chunks.
- Whisper: bumped beam_size 1→5 and added RO initial_prompt.
- Tests: 11 DAVE unit tests + 2 callback integration tests + contract test
with fork-version guard.
## Voice UX polish (d1bc77e)
- Killed the 3s "mă gândesc" filler (always collided with Claude p50 4-7s).
- Barge-in via `ttsq.clear()` at top of `on_segment_done`.
- DTX silence-flush poller (200ms tick) — Discord stops sending RTP packets
when silent, so the inline silence-check in sink.write() never fired for
trailing audio; background thread handles it.
- `EchoStreamingAudioSource.read()` non-blocking — old `get_frame(timeout=0.1)`
wrecked Discord's 20ms cadence and the client interpreted bursts as
stuttering (Marius heard "4 de minute" instead of full sentence).
- RO time expansion: 23:09 → "douăzeci și trei și nouă minute".
- Supertonic Unicode sanitize centralized in tools/tts.py.
- Whisper local_files_only=True — no HF metadata GET on each startup.
- Diagnostic logging through sink → VAD → Claude stream → TTS chain.
## Voice mode iteration (e589e48)
- `personality/VOICE_MODE.md` — voice-tailored system prompt (short, no
markdown, no abbreviations, time without seconds, distances in
"mii"/"milioane"); plumbed via build_system_prompt(voice_mode=True).
- Isolated voice session key `voice:<channel_id>` — voice doesn't share
context with text adapter on the same channel; auto-applied without
/clear ceremony. /clear drops both keys.
- Metric units + Romanian thousands (normalize.py): "384.000 km" →
"trei sute optzeci și patru de mii de kilometri" with feminine-correct
pluralization and "de" particle for ≥20.
- `/voice setvoice <M1-F5>` slash command with native autocomplete; swaps
live + persists voice.default_voice to config.json.
- Verbal voice change (src/voice/voice_commands.py + 29 tests) — "schimbă
vocea pe M5", "voce em cinci", with permissive substring fallback for
Whisper-mangled forms like "Mâcinci"=M5 and "unul cinci"=M5. Whisper
initial_prompt now lists voice vocabulary to bias STT toward clean
outputs.
- Fast barge-in: VAD ≥2 consecutive windows (~200ms) on Marius's user
while Echo has pending TTS frames → cut him off mid-sentence so user
doesn't wait the full silence + STT cycle. Acoustic echo bleed-through
still requires headphones (no AEC).
## Test suite
130 voice + router tests pass (test_voice_recv_dave, test_voice_session_cleanup,
test_voice_adapter_contract, test_voice_normalize, test_voice_commands,
test_router).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
