Echo
A voice is not delivered when it is made, but when another body receives it.
Echo is the speech-lifecycle Extension Domain. It carries an utterance from deliberate capture to actual reception without losing its origin, consent, or outcome. Audio admission is Partial, but no working Echo package or speech path ships. ADR 37 owns the accepted design; State of Work owns delivery truth.
The organs of resonance
An Echo deployment composes five roles independently. Audio is material, never a capability
family; stt, tts, and chat below are current v1 compatibility labels, not the Designed
general-service ABI.
| Role | Place in Echo |
|---|---|
| Ear | echo.transcribe@1 Animator; current v1 projects stt |
| Voice | echo.synthesize@1 Animator; current v1 projects tts |
| Audio-capable Mind | Audio-input or audio-output model.chat@1 provider; current v1 projects chat |
| Listener | Capture, push-to-talk, codecs, and optional voice activity |
| Mind | Ordinary reasoning provider; there is no separate voice intelligence |
First engine profile
The first planned local profile deliberately admits one inference engine:
| Concern | Initial choice | Boundary |
|---|---|---|
| Engine | audio.cpp | One pinned local runtime; support is family-specific, never arbitrary GGUF execution. |
| Ear candidate | Parakeet TDT 0.6B v3 | Multilingual STT candidate whose Slovak path must pass the bake. |
| Voice candidate | Supertonic 3 | Multilingual TTS candidate whose Slovak voices and OpenRAIL-M terms must pass the bake. |
| Connector | OpenAI audio batch dialect | Bounded transcription and speech routes only; runtime-native extras are not inferred. |
| Activation | Visible push-to-talk | No always-on capture and no wake-word dependency. |
| Interaction | Record and send | No claim of full-duplex speech or barge-in. |
This is designed direction, not delivered integration. The initial implementation should prove audio-to-text and text-to-audio before exposing audio.cpp voice conversion, cloning, alignment, VAD, diarization, codecs, or pipelines. Those functions become separately declared tool surfaces only after their model, inputs, outputs, custody, authority, and cancellation behavior are baked; the presence of a family in an upstream catalogue grants no LychD capability.
One connector may serve several compatible model profiles, but every profile keeps its own model identity, language facts, license, backend, resource measurements, and readiness. A language can be declared from upstream metadata without being verified by LychD. Routing uses an explicit request first, then ordered Principal or Persona preferences, then admitted detection; missing output-language support refuses rather than silently translating.
Watched alternatives
The first profile does not implement these paths:
| Candidate | Why it remains watched | Promotion condition |
|---|---|---|
| NeMo-Speech.cpp | Narrower NVIDIA speech runtime with promising realtime, endpointing, word-timing, and Riva surfaces. | Stable pinned release and dialect, lifecycle and cancellation receipts, and a multilingual realtime bake that materially beats the audio.cpp path. |
| whisper.cpp | Mature, portable Whisper-only STT fallback. | A required platform or recovery case that audio.cpp cannot satisfy. |
| Piper | Small CPU TTS fallback with separately licensed voices. | A measured low-resource or recovery need that Supertonic through audio.cpp cannot satisfy, plus runtime and voice-license admission. |
| openWakeWord | Host-side wake-word candidate, not a speech server. | An authorized phrase/model with acceptable license, reproducible training or acquisition, and measured false-accept and false-reject rates in the target language and environment. |
NeMo-Speech.cpp may later reuse the batch connector through an explicit dialect, while its live WebSocket still needs a Resonance Session adapter. Whisper, Piper, and openWakeWord are not installed merely as speculative redundancy. No watched candidate is an automatic fallback, and no failed local attempt silently changes engine, language, or egress.
The first bake
The smallest honest evaluation fixes one runtime revision and model digest, then records:
- cold start, readiness, clean shutdown, cancellation, CPU/GPU placement, RAM/VRAM peak, and isolation from a resident Mind;
- transcription accuracy, endpoint truncation, names, numbers, mixed-language speech, confidence, and provenance for each language claimed as verified;
- synthesis latency, first audio, intelligibility, pronunciation, numbers, dates, currency, addresses, text normalization, and playback completion for each output locale and voice;
- exact OpenAI-compatible fields and errors rather than compatibility by endpoint name; and
- license and consent receipts before any voice-reference or cloning experiment.
Slovak is the first operator-selected reference bake, not a Core hardcode. Other users may order different input and output preferences, but every routed language owes the same evidence.
The resonance path
Record-and-send is the first accepted profile:
visible bounded capture → immutable audio artifact → eligible Ear or audio-capable Mind → attributed transcript or native observation → ordinary Agent step → optional Voice → delivery evidence
Capture must be visible, bounded, and consented. Large bytes remain outside Graph checkpoints and queues.
Three distinctions keep the path honest:
- Current
ArtifactRefsupport is metadata only. The artifact-reference boundary proves neither a stored recording nor playable byte custody. - In the accepted design, Reliquary custody preserves source and derivation. A transcript is an attributed interpretation, not a replacement for its recording.
- Synthesized audio is a new artifact. Generation proves neither delivery nor playback; receipts must state what actually reached the listener.
Later half- and full-duplex profiles must preserve one monotonic speech timeline through barge-in, contested turn ownership, interruption, cancellation, and reconnect. A live socket cannot supply those semantics.
Invalid audio, insufficient authority, and policy-ineligible remote egress are refused. Only an otherwise eligible managed provider that is not warm may cause the requesting Run to enter ordinary Graph Stasis while Orchestrator drains and readies the affected Animators. Echo cannot silently fall back to a remote service, infer permanent microphone authority, inflate priority, or revoke another grant or lease.
Applications may use the voice without owning Echo
Echo owns speech contracts and lifecycle facts, not every work containing a voice. Riffmaw owns sung and otherwise musical performance; Language Edition owns casting, linguistic adaptation, timing, captions, and acceptance for one timed-language edition; Avatar may reference an exact eligible acoustic voice inside a presentation profile; and Broadcast owns canonical source words, editorial relation, and publication. Each application references exact Echo attempts and artifacts rather than copying its capture, synthesis, delivery, or playback chronology.
The Companion device boundary
A future Companion client may own device capture and playback as Echo's physical ear and mouth. A private Tether supplies reachability, not caller authority: Ward still authenticates and authorizes, while capture and consent remain visible. A remote Portal additionally requires eligible classification, egress policy, consent where required, and bounded cost. These are laws of the accepted design, not claims of delivered behavior.
Operationally, Echo preserves where speech began and whether the answer arrived.