# Glossary The research pages use terms from three fields that do not usually meet — speech recognition, human-computer interaction, and text-entry evaluation. This is the shortest useful definition of each, in the sense this project uses it. Where a term has a **stronger claim than it appears to**, that is said. Several of the disagreements on these pages are really disagreements about a definition. ## Recognition **ASR** — Automatic Speech Recognition. Audio in, text out. Everything else on this page is about what happens either side of that. **WER — word error rate.** Substitutions + insertions + deletions, divided by the number of reference words. The field's default metric, and **the source of a specific blind spot**: it weights every word equally. A wrong `their` and a wrong `rm -rf` count the same, which is why the [problem space](https://github.com/MSKazemi/yazses/blob/main/design/research/2026-08-15-problem-space.md) treats cost-weighted error as a separate, unmeasured thing. **Beam search / greedy decoding** — whether the decoder keeps several candidate transcriptions alive or commits to the best next token. Beams cost time and buy accuracy; on a CPU that trade is the whole latency story. **Hallucination** — a recogniser emitting fluent text for audio that contained no speech. Whisper is known for it on silence, which is why the silence gate exists *before* the decoder rather than after it. **Initial prompt** — text handed to the decoder as context so it biases toward your vocabulary. YazSes uses it for the personal dictionary and for the app name itself. **Hotword / contextual biasing** — the same goal at decode time rather than as a prompt. Stronger, and engine-specific; the reason `hotwords` is still unwired. ## Capture **VAD — voice activity detection.** Deciding which audio contains speech. YazSes's default is an energy gate: `mean(|samples|)` against a threshold. Cheap, and the thing most often misconfigured — below your voice and everything is discarded, above the room and nothing is. **Endpointing** — deciding *when you stopped speaking*. Distinct from VAD, which decides whether a moment is speech. Endpointing is what makes release feel instant or laggy. (Confusingly, YazSes's `EndpointConfig` is about this, not about network endpoints.) **Pre-speech padding** — a ring buffer of audio kept from *before* the key was pressed, prepended to the recording. Without it the first phoneme is clipped, because humans start speaking fractionally before they finish pressing. **Barge-in** — speaking while the system is talking back, and having it stop. ## Speaker **Diarization** — "who spoke when". Segmenting audio by speaker without necessarily knowing who they are. **Speaker embedding / d-vector** — a fixed-length vector representing a voice, so two utterances can be compared for sameness. **Biometric**: this is why ADR-019 forbids exporting them under any consent regime. **ECAPA-TDNN** — the embedding model YazSes uses by default. Accurate on utterances of a second or more, and **unreliable below about half a second**, which is exactly why the Cocktail Filter false-rejects the user's own voice and stays off. **Enrolment** — recording a known speaker once so later audio can be matched to them. Opt-in and consent-gated here, never automatic. ## Language **Code-switching** — changing language *within* an utterance, not between sessions. The [open question](https://github.com/MSKazemi/yazses/issues/258) is whether this, rather than accent, is the dominant failure for bilingual users. **LID — language identification.** Deciding which language is being spoken. Per-utterance LID is standard; per-*span* LID is what code-switching needs and what makes it hard. **ITN — inverse text normalisation.** Turning spoken forms into written ones: "twenty twenty six" → "2026", "john dot doe at gmail" → an address. ## Commands and structure **Grammar (Tier 1)** — the regex classifier that decides whether an utterance is dictation or a command. **Anchored at both ends** in this codebase, because ordinary words like "undo" and "cancel" are also commands, and a loose match swallows real speech. **GBNF** — the grammar format llama.cpp uses to force a model's output into a schema. How a local model can be made to emit valid tool calls rather than plausible-looking ones. **MCP — Model Context Protocol.** The standard interface between an agent and the tools it can call. [ADR-020](https://github.com/MSKazemi/yazses/blob/main/design/adr/adr-020-agent-protocols.md) covers which direction of it YazSes should support and why. **AT-SPI** — the Linux accessibility bus. How an application can be asked what has focus and whether it accepts text. The basis of the no-text-target guard, and unavailable on Wayland without it, which is why several capabilities are X11-only. ## Evaluation **Corrected / uncorrected error rate** — errors the user fixed versus errors that survived into the final text. Reporting only the second flatters a system that made the user work hard; reporting only the first hides what shipped. Text-entry research reports both. **MSD error rate** — minimum string distance between what was presented and what was produced, normalised. With the two above, the standard set from Soukoreff & MacKenzie. **WPM — words per minute**, conventionally with a "word" defined as five characters so that languages and scripts compare. The 153-vs-52 figure on these pages uses this. **Held-out** — evaluated on data the model was not tuned on. The bar any personal-adaptation claim has to clear, and the reason that work stays unbuilt rather than merely unfinished. **Preregistration** — writing down the hypothesis and analysis *before* collecting data, so a result cannot be reverse-engineered from what the data happened to show. ## This project's own vocabulary **Wired / unwired** — whether a capability in the registry is reachable from a real entry point. An unwired capability may be fully implemented and tested; it simply has no door. Roughly 62 of 144 are unwired, and the count is [enforced by a test](https://github.com/MSKazemi/yazses/blob/main/tests/test_feature_wiring_honesty.py) so the catalogue cannot overstate itself. **Hold-to-talk** — recording only while a key is physically held. The alternative, a toggle, is what produces the "was it listening?" class of problem this project's design avoids. **Egress path** — any code that can open an outbound connection, classified *fetch* (data comes in) or *send* (user content goes out). There are two of the second kind, and [ADR-019](https://github.com/MSKazemi/yazses/blob/main/design/adr/adr-019-egress-inventory-and-escalation.md) keeps it that way with a test.