| Term | Meaning here |
|---|
| A2P / 10DLC | Application-to-person messaging over ten-digit long-code numbers. Carriers require a registered brand and campaign with a verifiable opt-in before business SMS deliver reliably. |
| Backchannel | A short acknowledgment ('mm-hm', 'let me check') that keeps a conversation feeling alive. In the director roadmap these are pre-synthesized clips the talker plays while the brain works. |
| BAA | Business Associate Agreement: the HIPAA contract that must cover a vendor's specific service before PHI may touch it. See the compliance page for the full chain. |
| Barge-in | The caller talking over Elle and being honored: generation, synthesis, and playback all cancel, and her memory is truncated to what was actually heard. |
| BM25 | The classic lexical ranking function behind the knowledge search; retrieval with no embeddings and no vector database. |
| Cache breakpoint | A cache_control marker on a prompt block telling the API 'everything up to here may be reused.' One rides the static persona; another rides the conversation's last message. |
| Cascade | The architecture family where STT, LLM, and TTS are separate components passing text: maximal control, timer-based turn-taking. Elle's production stack is an advanced cascade. |
| Code-switching | Changing language mid-conversation, even mid-sentence. Nova-3's multi mode transcribes it in one stream, tagging each word's language. |
| Containment | The fraction of calls the AI resolves end to end with no human. The single most valuable number in the call-center economics: one point ≈ $55K/month at modeled volume. |
| CTA (call to action) | In A2P review, the public place where a consumer learns how to opt in. Reviewers must be able to find and verify it; ask.html carries Elle's. |
| Director | The roadmap architecture: hearing and turn-taking fused into one duplex model that answers routine turns itself and escalates hard ones to the brain, whose sentences stream straight to the house TTS. |
| Duplex model | A model that listens and speaks simultaneously rather than in strict turns, making turn-taking judgment part of the model instead of a timer. |
| Endpointing | The STT's decision that a speaker has finished, from trailing silence. Tunable; the cascade's one native compromise, and what the director removes. |
| Escalation | The talker handing a hard turn to the inference brain via a tool call, invisibly to the caller. The delegation-leak eval exists to keep it invisible. |
| Fusion boundary | Which of the four layers (hearing, judgment, reasoning, speech) are merged into one model. Every voice architecture is a choice of boundary; the trade-offs page prices them. |
| Greeting gate | The bridge-side lock that keeps barge-in disabled until every frame of the opening message has been sent to the caller. |
| Half-cascade | Industry term for audio-in, text-middle, synthesized-out architectures; the director is a full-fledged one with an escalation tier. |
| μ-law (G.711) | The 8 kHz telephone audio codec Twilio speaks. Both speech legs run it natively, so no transcoding happens anywhere: 8 bytes per millisecond, 160 bytes per 20 ms frame. |
| Media stream | Twilio's bidirectional WebSocket carrying live call audio; guarded by the HMAC media token. |
| Pairing code | The short code Elle speaks that joins a browser to the live call for transcripts, visuals, and typed input. |
| PHI | Protected Health Information under HIPAA. Elle handles none; the compliance page documents the path a deployment that does would take. |
| Prompt cache | Provider-side reuse of an identical prompt prefix at ~10% of input price. Three layers of it cut the LLM line ~85%. |
| Silent-call guard | The timer that hangs up a call whose caller has said nothing (robocall probes); cleared by the first real utterance, typed input included. |
| Smart turn / semantic VAD | Vendor terms for end-of-turn detection smarter than a silence timer; the shallow cousin of what the director fuses. |
| Speech-to-speech (fused) | One model consuming and producing audio directly. Most natural turn-taking, but paraphrased output, voice lock-in, and 5-10x cost; Elle's demo mode. |
| Static / dynamic split | The prompt discipline making caching work across calls: byte-identical persona first, per-call context after, so the shared prefix never breaks. |
| Talker / thinker | Informal names for the director's two halves: the cheap fused model that handles most turns, and the full LLM it escalates to. |
| turnSeq | The integer at the heart of cancellation: every turn captures it, barge-in increments it, and any stage whose copy is stale stops at its next checkpoint. |
| TwiML | Twilio's XML instruction language; the /voice-realtime webhook answers with TwiML that opens the media stream and embeds the media token. |
| Utterance | One complete caller turn as assembled by the STT from final transcripts, ended by endpointing or its slower fallback. |
| VAD | Voice activity detection: 'sound is happening.' Deliberately NOT trusted for barge-in, which requires recognized words. |
| Verbatim delivery | Speaking a script exactly as written; guaranteed by the cascade's text bottleneck and impossible to guarantee in fused speech-to-speech. |
| WAL | SQLite's write-ahead log mode, letting live-call writes and admin reads coexist without blocking. |
| Warm socket | A connection opened before it is needed (TTS during LLM thinking) so the handshake never delays the first audible word. |
| ZDR | Zero data retention: the API configuration OpenAI requires alongside its BAA before PHI may touch the Realtime API. |