Glossary

Plain-English definitions of transcription and knowledge-base terms.

Agent memory: The persistent store an AI agent uses to retain and recall information across an ongoing task or conversation — distinct from one-off retrieval (RAG).
Air-gapped: A deployment with no connection to external networks — physically or logically isolated from the internet — so data cannot leave and remote access cannot reach in. It is the strictest form of on-premises operation.
Attorney-client privilege: The legal doctrine that protects confidential communications between a lawyer and client from forced disclosure — protection that extends to recordings, notes and transcripts of those communications.
Automatic speech recognition (ASR): The technology that converts spoken audio into text automatically. Modern ASR is powered by deep-learning models such as Whisper.
Business Associate Agreement (BAA): A HIPAA-required contract between a healthcare entity and a vendor that handles protected health information (PHI), defining how the vendor safeguards that data.
Chunking: Splitting a long document or transcript into smaller pieces ("chunks") so they can be embedded and retrieved individually by a search or RAG system.
Data residency: The requirement that data be stored and processed within a chosen country or jurisdiction — which decides whether a recording can go to a given cloud region or must stay on local infrastructure.
Embeddings: Numeric vector representations of text (or other data) that capture meaning, so that similar content sits close together in vector space.
faster-whisper: A reimplementation of OpenAI's Whisper model on the CTranslate2 inference engine, running several times faster with lower memory use than the reference code. It is a common engine inside self-hosted transcription services.
FERPA: The US Family Educational Rights and Privacy Act, which protects the privacy of student education records — including lecture, advising and interview recordings — at schools and universities that receive federal funding.
GDPR: The EU General Data Protection Regulation, which governs how the personal data of people in the EU/EEA is processed — including recordings of identifiable individuals — and requires a lawful basis, data minimisation and care with cross-border transfers.
GLBA: The US Gramm-Leach-Bliley Act, which requires financial institutions to protect customers' non-public personal financial information — including recorded calls and their transcripts — under its Safeguards Rule.
Hallucination (ASR): In speech recognition, a hallucination is transcript text the model outputs for words that were never spoken — a known Whisper failure mode on silence, music, or background noise.
HIPAA: The US Health Insurance Portability and Accountability Act, which sets national rules for protecting patient health information (PHI) — including how it may be recorded, transcribed, stored and shared by healthcare providers and their vendors.
Knowledge base: A structured, queryable store of information; in this context, one built from your own transcribed audio so people or AI agents can search and cite it.
Model Context Protocol (MCP): An open standard that lets AI agents connect to external tools and data sources through a common interface — including memory and knowledge stores.
On-premises (on-prem): Software that runs on infrastructure you control — your own servers or workstations — rather than a vendor's cloud. Data never leaves your environment.
pgvector: An open-source extension that adds vector similarity search to PostgreSQL, so a normal Postgres database can store embeddings and power semantic search / RAG.
Real-time transcription: Transcribing audio as it is spoken, with very low latency, rather than processing a complete recording after the fact (batch transcription).
Retrieval-augmented generation (RAG): A technique where an LLM retrieves relevant documents from a knowledge store (often a vector database) and uses them to ground its answers.
Semantic search: Searching by meaning rather than exact keywords, by comparing the vector embedding of a query against stored embeddings and returning the closest matches.
SOC 2: A voluntary audit framework from the AICPA that reports on how a service organisation controls customer data across five trust criteria: security, availability, processing integrity, confidentiality and privacy.
Speaker diarization: The process of partitioning an audio recording by speaker — determining "who spoke when" — and labelling each segment (Speaker 1, Speaker 2, …).
Speaker embedding: A numeric vector that represents a speaker's voice characteristics, used to tell speakers apart and to recognise the same speaker across recordings.
Subtitles (SRT / VTT): Time-coded text tracks that display transcribed speech on screen. SRT and WebVTT are the two most common subtitle file formats.
Transcription: The conversion of spoken audio into written text, usually by an automatic speech recognition (ASR) model such as Whisper.
Vector database: A database that stores data as high-dimensional vectors (embeddings) and retrieves by semantic similarity rather than exact keywords.
Voice activity detection (VAD): Voice activity detection (VAD) is a preprocessing step that identifies which parts of an audio stream contain human speech versus silence, music, or noise, so a transcription model only processes the spans that carry words.
WhisperX: WhisperX is an open-source transcription library that runs OpenAI's Whisper model and adds a forced-alignment pass for word-level timestamps plus an optional speaker-diarization hook, producing timing-accurate transcripts with per-word start and end times.
Word error rate (WER): A standard accuracy metric for transcription: the percentage of words the system gets wrong (substitutions, insertions and deletions) versus a reference.
Word-level timestamps: Word-level timestamps are per-word start and end times produced by forced alignment (for example, WhisperX), pinpointing exactly when each word is spoken rather than just when a segment begins and ends.