Custom vocabulary and technical acronyms

Every engineering team and professional firm relies on unique terminology: internal project codenames, technical jargon, medical nomenclature, and foreign surnames. Out of the box, generic speech models can mishear these phrases. Oyma allows you to supply custom vocabulary prompts to steer Whisper’s decoder.

How initial prompt biasing works

OpenAI’s Whisper architecture accepts an initial text prompt that establishes linguistic context for speech decoding. When you register custom terms in Oyma, the speech engine feeds these phrases into the decoder prior to processing the first audio frame:

[Custom Terms: "ScreenCaptureKit, FTS5, Ollama, CommonMark, Wikilinks"]
                               │
                               ▼
┌─────────────────────────────────────────────────────────────┐
│                 Whisper Decoder Attention                   │
│   (Biased to recognize specialized terminology accurately)   │
└─────────────────────────────────────────────────────────────┘

Adding custom vocabulary

1. Open Settings › Transcription › Vocabulary.
2. Enter terms as a comma-separated list or newline-separated entries:

   ScreenCaptureKit, FTS5, Ollama, SQLite, CommonMark, APFS,
   Kubernetes, Datadog, Prometheus, GraphQL, WebRTC

3. Click Save Vocabulary.

Best practices for vocabulary tuning

• Capitalization Matters: Write acronyms in uppercase (OAuth, SLA, CI/CD) so Whisper outputs standard casing.
• Compound Technical Terms: Separate multi-word phrases with standard spacing (PostgreSQL, Tailwind CSS).
• Phonetic Hints for Foreign Names: If a colleague’s surname is frequently misspelled, include it in your vocabulary roster alongside common phonetics.

Related guides

• Model sizing in Transcription models.
• Speech benchmarks in Whisper benchmarks on Apple silicon.
• Audio batch processing in Batch audio and video transcription.