ADR 0281: Host speech capability

• Status: Accepted for implementation (amended by ADR 0291)
• Date: 2026-09-17
• Deciders: PI-Desktop core
• Related: ADR 0257 · 03-runtime/20-speech

Context

Chat models, image generation, transcription, and speech synthesis are different jobs. @earendil-works/pi-ai has no TTS/ASR surface, and Whisper / MIMO TTS must not appear in the chat model picker. Local OpenAI-Audio-compatible servers (Speaches, whisper.cpp, LocalAI) should work through an existing openai_compatible provider.

Decision

1. The host owns a speech capability independent of chat: transcribe(audio) → text and synthesize(text) → audio.
2. Bindings live on optional AppSettings.speech (no schema bump). Each role names an existing provider, a model id, and an open protocol id.
3. Built-in protocols: openai_audio (REST /audio/transcriptions and /audio/speech) and openai_chat_audio (chat completions audio field; MIMO mimo-v2.5-tts). New vendors add an adapter, not a new IPC channel.
4. Plugins may register a protocol with pi.speech.registerAdapter under high-risk speech.adapter.register. Handles stay in the guest; HTTP plans are executed by the host with the bound provider's key and must stay on that origin. Built-in protocol ids are reserved.
5. v1 product entry is the host API only: the speech/* IPC and plugin adapters. No Settings card and no Composer transcription / draft-speech control exists (ADR 0291 withdrew both). Audio bytes never enter the renderer (path in, scratch out).
6. Out of scope: microphone / pi.audio device backend, Realtime, agent transcribe/speak tools, audio as LLM content blocks, changing pi-ai.

Consequences

An unconfigured role fails SPEECH_NOT_CONFIGURED. Provider deletion makes the binding fail NOT_FOUND. Plugin unload drops its protocols.