STT v3 — route hermes-agent to whisper.service HTTP + cloud OpenAI primary

[MKheru]

Summary

Migrates the STT path used by hermes-agent for inbound voice messages from faster-whisper base (in-process, ~39M params) to:

1. Cloud primary: OpenAI gpt-4o-mini-transcribe (sub-second, $0.003/min, no hallucinations on 4-audio bench)
2. Local fallback: whisper.service HTTP at http://127.0.0.1:8765/inference (whisper.cpp server already running with ggml-large-v3-turbo)

The local fallback closes a longstanding mismatch: whisper.service was already deployed with the large-v3-turbo model but hermes-agent silently loaded the much smaller base model in-process via faster-whisper.

What's in the diff

tools/transcription_tools.py

• New _transcribe_via_whisper_service(file_path, endpoint, language) function. Pure stdlib urllib (no requests dep), manually-constructed multipart form-data, OpenAI-compatible response shape. Returns same dict as legacy _transcribe_local().
• _transcribe_local() now tries the HTTP endpoint first (when stt.local.endpoint is configured), falls back to in-process faster-whisper on any HTTP failure.

tools/voice_io.py (newly tracked)

• WS16 voice loop client. Talks to whisper.service (STT) and tts.service (TTS) via HTTP. No subprocess (SECCOMP-compatible).

uv.lock

• Auto-updated by uv during the session.

Validation

Bench performed by Claude on 4 real audios (Khéri's Discord vocals, archived at ~/.hermes/cache/audio/):

Backend	Latency	Hallucinations	Cost
whisper-1 (cloud)	1.2–4.6s	1/4 totalement halluciné ("Merci Philip..." + leak chinois)	$0.006/min
gpt-4o-mini-transcribe (cloud)	0.7–2.3s	0/4	$0.003/min
whisper.service large-v3-turbo (local)	23–47s (CPU only)	0/4	$0

gpt-4o-mini-transcribe retained as primary. whisper-1 disqualified on hallucination rate. whisper.service retained as fallback (slow but free + already deployed).

User-facing config

Requires ~/.hermes/config.yaml to be updated to set:

stt:
  provider: openai
  local:
    endpoint: http://127.0.0.1:8765/inference
  openai:
    model: gpt-4o-mini-transcribe

Config not in this PR (it's user runtime, not in repo). Applied to prod separately.

Test plan

• Smoke test: _transcribe_via_whisper_service against archived OGG → success + correct French
• Live test: Discord voice message transcribed end-to-end via new route
• Native voice reply path tested (depends on discord.py json import fix in companion PR)