Make speech (STT/TTS) provider-agnostic (SPEECH_PROVIDER, OpenAI-compatible backends)

[Averroeskw]

Why

Speech is the last assistant-critical service that's hard-wired to specific vendors: TTS requires an Azure Speech key (MSFT_SPEECH_KEY) and STT a Groq key, with no way to point either at a different backend. Same single-vendor situation as search/weather before #8 — anyone without those exact keys (or running fully self-hosted) has no voice pipeline at all.

What

Speech is now provider-selectable via SPEECH_PROVIDER (server/src/services/speech/):

• azure (default) — the stock pipeline, byte-for-byte unchanged: Azure Speech synthesis + Groq-hosted Whisper recognition. (Naming note: STT was already Groq, not Azure; "azure" here just means "the stock stack".)
• openai — any OpenAI-compatible audio API via POST /audio/transcriptions (multipart) and POST /audio/speech: OpenAI itself, a LiteLLM proxy, self-hosted faster-whisper / Speaches, openedai-speech, etc. Configured with OPENAI_SPEECH_BASE, OPENAI_SPEECH_KEY (optional for keyless self-hosted servers), OPENAI_STT_MODEL (default whisper-1), OPENAI_TTS_MODEL (default tts-1), OPENAI_TTS_VOICE (default alloy).

If SPEECH_PROVIDER is unset, openai is auto-selected only when OPENAI_SPEECH_BASE is configured — otherwise behavior is exactly as today.

Audio formats (the honest part)

• Azure TTS returns Ogg16Khz16BitMonoOpus, and the downstream pipeline depends on that: changeVolume / addBackgroundAudio force inputFormat("ogg") before the audio reaches the device.
• The OpenAI provider therefore requests response_format: "opus" (Ogg/Opus, 48 kHz) and transcodes to Ogg 16 kHz mono with the ffmpeg toolchain the server already uses (new transcodeToOgg helper in services/audio.ts), so everything downstream sees the same profile regardless of provider. One extra ffmpeg pass per response; no client/device changes.
• recognize() interface unchanged (Buffer in → transcript out); the device's Ogg capture buffer is uploaded as audio.ogg, which Whisper-style endpoints accept natively. The silence gate (WHISPER_MIN_DB) stays provider-independent.
• The dashboard voice picker maps to Azure voice names, so it only applies to the azure provider; the openai provider takes its voice from OPENAI_TTS_VOICE. speed carries over cleanly (both are 1.0-based multipliers, clamped to OpenAI's 0.25–4 range). The translate feature's language-identifying STT (Google Speech v2) is a separate code path and is intentionally left untouched.

Compatibility & verification

• Azure/Groq deployments: zero behavior change (default provider, code path untouched).
• .env.template documents the new block and fixes the stale SPEECH_KEY name — the code reads MSFT_SPEECH_KEY.
• tsc --noEmit clean; new files lint clean.
• Smoke-tested both directions against an OpenAI-compatible stub: multipart STT (auth header, model field, filename) and TTS request shape verified, output confirmed Ogg / 16 kHz / mono via ffprobe and accepted by the existing changeVolume path.

🤖 Generated with Claude Code

[Averroeskw]

Closing for now — pulling this back until I've verified the full stack end-to-end on the physical device (interposer inbound). Will reopen once it's hardware-verified. The branch lives on in my fork history if anyone needs the diff in the meantime.