Reference

Glossary.

Plain-language definitions for the technical terms across Bonfiyah and the meeting-transcription category. If a word in the docs sounded jargony, it's defined here.

AI Project Context: A 250–450 word executive briefing synthesised across every recording in a Bonfiyah project. Cached against a content signature so it stays fresh as recordings are added or edited. The cross-recording brief that lets you re-enter a project after weeks away in 90 seconds. · Deep-dive →
the upstream transcription provider: Bonfiyah's full-transcript engine. A cloud transcription API that produces speaker-diarized transcripts with high accuracy across 30+ languages. Audio is sent to the upstream transcription provider only when full transcription is enabled; we opt out of the upstream transcription provider's model-training program, and each transcript is deleted from the provider immediately after processing (their retention floor is a 1-hour TTL), so your content is not used as training data. Live captioning runs on-device via Apple Speech and does not use the upstream transcription provider.
Apple Speech: iOS's built-in on-device speech recognition framework (SFSpeechRecognizer). Bonfiyah uses it for live captioning during recording — sub-second latency, fully on-device, no audio leaves the iPhone. Less accurate than an upstream transcription provider for the full transcript but ideal for real-time UX.
Compatibility Analysis: A pairwise scorecard combining four research-grounded frameworks (Attachment Theory, Big Five / OCEAN, Gottman Four Horsemen, Thomas-Kilmann Conflict Modes) computed on a recorded conversation between two people. Refuses to run without explicit two-party consent — enforced in code, no override branch. · Deep-dive →
Cohort (in cross-recording Voice ID): The set of devices linked to a single Bonfiyah user via iCloud — iPhone, iPad, Mac via Catalyst, and any historical recordings on those devices. Speaker library is scoped to the cohort. Voice signatures match within the cohort, never across cohorts.
Diarization: The process of labelling each utterance in a recording with a speaker identity. Bonfiyah does first-pass diarization via the upstream transcription provider, then a per-utterance an on-device correction pass pass that splits hybrid clusters automatically. Result: 'Speaker A said X, Speaker B said Y' rather than a wall of text with no attribution.
Dynamic Island: iPhone 14 Pro and later have a pill-shaped cutout at the top of the screen that hosts Live Activities. Bonfiyah's recording state lives there when the app isn't frontmost — elapsed time, current speaker, one-tap pause. Tap to expand, swipe to dismiss.
voice recognition: Emphasised Channel Attention, Propagation, and Aggregation — Time-Delay Neural Network. The speaker-verification architecture Bonfiyah ships in the app bundle (compact, runs on-device via Apple's on-device AI on your device). Generates compact voice embeddings used to identify the same speaker across recordings and to split the upstream transcription provider's hybrid diarization clusters. · Deep-dive →
Email Intelligence: Bonfiyah's term for the deep-linked section headings in emailed exports. Every emailed AI Summary, Speaker Insights, and Team Dynamics PDF carries tappable headings — tap 'Decisions' or a speaker name in your inbox, Bonfiyah opens directly to that section. The links resolve to a public web view if the recipient doesn't have Bonfiyah, never to a broken page. · Deep-dive →
FAQPage schema: Schema.org structured data type that exposes a page's question/answer pairs to search engines and AI answer engines. Bonfiyah's feature pages emit FAQPage JSON-LD so Google can show rich-result snippets and ChatGPT/Claude/Perplexity can extract Q&A directly. Not user-facing — an SEO/GEO/AEO substrate.
Hybrid cluster (diarization): When a diarization engine merges two speakers' utterances into a single label because the voices are acoustically similar. Bonfiyah's per-utterance an on-device correction pass pass detects and splits these automatically; in our internal evaluation it corrects ~73% of the upstream transcription provider hybrid errors before a user ever sees them. The remaining 27% are catchable manually via Resplit Voice Matching.
Live Activity: An iOS feature that lets apps display real-time information on the lock screen and in the Dynamic Island. Bonfiyah uses it to show recording state — elapsed time, peak-volume meter, current speaker, one-tap pause — without requiring you to unlock the phone or open the app.
Local notification: An iOS notification scheduled by the device, not pushed via Apple Push Notification Service (APNs). Bonfiyah's Proactive Notifications are entirely local — the candidate feed is delivered as a list to the iPhone, where iOS schedules each as a local notification. The notification body content is never visible to APNs servers, never logged in Bonfiyah's infrastructure, and never visible to a third-party push provider. · Deep-dive →
Pre-Brief: A one-page brief delivered 30 minutes before a meeting, generated automatically from prior recordings with the same person. Pulls open commitments, recurring themes, and unresolved decisions. Reads your iPhone Calendar (read-only) to know when the meeting is. · Deep-dive →
Proactive Notifications: Bonfiyah's term for quote-driven notifications that surface specific events from your library — 'Sarah Hendron: 4 days overdue — "I'll send the Q4 deck by end of week."' Three tiers (Critical, Standard, Insights), daily caps, quiet-hours respect, and tap-to-exact-moment deep links. The first conversation app that reaches out to the user instead of waiting to be opened. · Deep-dive →
Promise Tracker: Pro AI feature that auto-extracts every commitment from every recording. Each commitment becomes a card with the speaker, deadline, and source quote. Tracks who promised what, by when, and whether they kept it. Cross-recording: a promise made Tuesday is still tracked Friday. · Deep-dive →
Story Mode: A narrative recap of a recording — characters, decisions, the moments that mattered. Distinct from AI Summary (descriptive bullets); Story Mode is prose. Story Episodes group recurring topics across recordings into narrative arcs. · Deep-dive →
Team Dynamics: A 9-box matrix placing N selected speakers across cohesion × drive (or alternative configurable axis pairs), with team cohesion score, swap recommendations, and coverage-gap analysis. Built from existing recordings — no separate dataset to assemble. · Deep-dive →
Truth Layer: Pro AI feature that catches when something said in a recording today contradicts something said in a prior recording weeks ago. Surfaces the contradiction with both source quotes and timestamps, before the next meeting where the topic comes up. · Deep-dive →
Two-party consent: A legal recording requirement in twelve U.S. states (California, Florida, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire, Pennsylvania, Washington, plus Connecticut and Oregon under specific conditions): every participant in a recorded conversation must be informed that the recording is happening. Bonfiyah ships built-in two-party consent management in every tier — verbal-consent capture and a defensible audit log so the on-record artifact each state requires actually exists. · Deep-dive →
Voice ID (cross-recording): The layer that makes 'Sarah from Tuesday's meeting is the same Sarah from yesterday's meeting' true across iPhone, iPad, Mac, and iCloud. Cohort-aware identity matching at the library level + per-utterance an on-device correction pass at the recording level. Required for Promise Tracker, People Memory, Compatibility, and Team Dynamics to be coherent. · Deep-dive →
WhisperKit: An on-device port of OpenAI's Whisper model, optimised for the iPhone's Neural Engine. Bonfiyah does not use WhisperKit today; live captioning runs on Apple Speech (SFSpeechRecognizer) and the full transcript runs on an upstream transcription provider.

More technical depth

Bonfiyah's engineering posts cover the prompt-caching strategy, the an on-device correction pass pass, and the architectural reasons for the privacy commitments. About once a week.

No spam. We use ConvertKit. See our privacy policy.