Term
Keyword Glossary (STT)
A keyword glossary for Speech-to-Text is a short, curated list of project-specific terms, brand and product names handed to an STT model as a context hint before transcription — typically through Whisper's `initial_prompt`.
Keyword glossary — explained in more detail
Standard STT models such as Whisper are trained on general web audio and know nothing about domain-specific vocabulary. A keyword glossary is the simplest countermeasure: a list of the ten to twenty most frequent proper names, tools, model and brand names that appear in your workflow. The list is passed to the model per call as a hint (Whisper: initial_prompt). The glossary does not change the model weights — it biases decoding towards known token sequences, so “DEVI Runner” gets recognised consistently instead of ending up as “Debbie Runner”. The leverage is high, the effort minimal: no training, no GPU, no infrastructure.
Example / Practical context
The most effective way to build a keyword glossary is data-driven: collect STT sessions, log every transcription error (e.g. by scoring each session into a CSV with an AI rater), and after ~90 sessions distil the most frequent errors into a list. Whisper-small tolerates roughly 15 terms before general recognition degrades; Whisper-medium leans toward 20–25. Example glossary from a coding workflow: DEVI Runner, boostN, Sonnet, Haiku, Heartbeat, foot pedal, transcription, Claude, CLI, JWT, commit, branch, node, worktree.
Distinction from related concepts
Custom vocabulary in commercial APIs (Deepgram, AssemblyAI) is conceptually the same — a project-specific term list, often with boost weights — but cast into the respective API rather than passed as a generic prompt. Fine-tuning the STT model on domain data is the next level: more precise, but more expensive and requiring a training pipeline. A post-processing dictionary corrects Whisper output after the fact (regex- or LLM-based) and addresses different error classes — such as word splits and hallucinations that a hint cannot fix.
Entdecke mehr
Headless without an API bill — how do you reach the best AI models for automation in 2026?
Provider comparison mid-2026: who has a headless mode, whose subscription still covers it — and why BYOK is the most stable foundation.
GlossarEU AI Act (KI-Verordnung)
Regulation (EU) 2024/1689 is the EU's first comprehensive law on artificial intelligence. It regulates AI systems by risk in four tiers, from prohibited practices to minimal risk, and sets obligations for providers and deployers.
LexikonFunction Calling / Tool Use
How an LLM uses tools: define a tool as a schema, the model picks the function and arguments, the result returns to the chat — the basis of every agent.