The short answer
Custom vocabulary means passing a list of words with each request so the recogniser favours them when the audio is close: Phonon-2 calls it hotwords (up to 25 terms), Deepgram calls it keyterm prompting (up to 500 tokens), AssemblyAI calls it keyterms_prompt (up to 1,000 words or phrases on Universal-3.5 Pro), and OpenAI and Whisper use a text prompt. The list biases recognition; it does not retrain the model.
Use it for names and terms the model has rarely heard. Keep the list short and specific to the request, because a long list of common words gives the model more ways to insert a term that was not said.
How each system exposes it
Syntax, limits and add-on prices from each provider's documentation and pricing page, read 2026-11-13.
| System | Feature | How to pass it | Limit | Extra cost |
|---|---|---|---|---|
| Phonon-2 (fermion, self-hosted) | hotwords | --hotwords "Ada, Quillon" or @names.txt; server field hotwords, or the OpenAI prompt field | 25 terms | none (runs on your hardware) |
| Deepgram Nova-3, Flux | keyterm prompting | keyterm= query parameter, repeated once per term | 500 tokens per request | $0.0013 per minute |
| Deepgram Nova-2 and older | keywords | KEYWORD:INTENSIFIER | see Deepgram docs | see Deepgram pricing |
| AssemblyAI Universal-3.5 Pro | keyterms_prompt | JSON list of strings | 1,000 words or phrases, 6 words per phrase | $0.05 per hour |
| AssemblyAI Universal-2 | keyterms prompting | JSON list of strings | 200 terms | $0.05 per hour |
| OpenAI gpt-transcribe | prompt, keywords | request fields | the model's length limit | no add-on listed |
| OpenAI whisper-1, Groq Whisper | prompt | request field | 224 tokens | no add-on listed |
| openai-whisper (open source) | initial_prompt | Python argument or --initial_prompt | text for the first window | none |
| faster-whisper (open source) | hotwords, initial_prompt | Python arguments | not stated | none |
How biasing works
A speech recogniser builds the transcript one piece at a time, scoring candidate word pieces against the audio. Hotword biasing adds a bonus to pieces that continue a listed term, so when the audio is ambiguous between 'Quill on' and 'Quillon', the listed spelling wins. When the audio clearly says something else, the bonus is not enough and the listed word stays out.
Prompting works differently. Models with a language-model decoder (Whisper, gpt-transcribe, Universal-3.5 Pro) read the prompt as preceding context, which makes the listed spellings more likely in the same way a sentence makes its next word more likely. AssemblyAI's docs say its contextual prompt should describe the audio, not give instructions, and keyterms carry the term list.
Deepgram's docs warn that commas, semicolons or weights inside a keyterm value are accepted without an error but boost nothing: pass one term per keyterm parameter.
Hotwords with Phonon-2
Phonon-2 takes up to 25 names and terms per request on every engine (Apple silicon, CPU on Linux, Windows and macOS, and the CUDA image). --hotword-lambda sets how hard the listed words pull (default 2.0). Without hotwords the transcript is unchanged, and hotwords sent with one server request do not affect other requests. Hotwords are not yet available on Phonon-2.5.
fermion transcribe phonon-2 call.wav --hotwords "Ada, Quillon, Okonkwo"
fermion transcribe phonon-2 call.wav --hotwords @names.txt
# Python
import fermion
speech = fermion.load_speech("phonon-2")
text, decode_s, audio_s = speech.transcribe("call.wav", hotwords=["Ada", "Quillon", "Okonkwo"])
# Server: an explicit field, or the OpenAI prompt field read as a vocabulary list
curl -s http://127.0.0.1:8000/v1/audio/transcriptions -F file=@call.wav -F model=phonon-2 -F "hotwords=Ada, Quillon, Okonkwo"Questions
- What is hotword biasing in speech recognition?
- It is a way to make a recogniser favour a list of words, such as names or product terms, when the audio is close to them. The model is not retrained; the list applies to the request that sends it.
- How many custom words can I add?
- It depends on the system: Phonon-2 takes 25 hotwords, Deepgram keyterm prompting allows 500 tokens per request, AssemblyAI Universal-3.5 Pro takes up to 1,000 words or phrases, and OpenAI whisper-1 prompts are limited to 224 tokens.
- Does custom vocabulary cost extra?
- On some APIs: Deepgram lists keyterm prompting at $0.0013 per minute and AssemblyAI at $0.05 per hour for pre-recorded audio. Self-hosted models such as Phonon-2 or faster-whisper have no extra charge.
- Can a hotword appear in the transcript when it was not said?
- It can on systems that bias too hard, which is why lists should stay short and specific. Phonon-2's documentation states that a listed word that was not said stays out of the transcript.
- How do I add custom vocabulary to Whisper?
- Pass the terms in a prompt: initial_prompt in the open-source openai-whisper package, prompt on the OpenAI and Groq APIs (224 tokens), or hotwords in faster-whisper.
- Does Phonon-2.5 support hotwords?
- Not yet. Hotwords are available on Phonon-2 today through the CLI, the Python API and the server.
Related
Sources and method
- Phonon-2 hotwords: /docs/speech/ and /docs/server-reference/ (fermion-research package documentation), read 2026-11-13.
- Deepgram keyterm prompting: developers.deepgram.com/docs/keyterm; add-on price: deepgram.com/pricing; read 2026-11-13.
- AssemblyAI keyterms and prompt: assemblyai.com/docs/pre-recorded-audio/universal-3-5-pro/prompting; add-on prices and limits: assemblyai.com/pricing; read 2026-11-13.
- OpenAI prompt and keywords: developers.openai.com/api/docs/guides/speech-to-text, read 2026-11-13. Groq prompt limit: console.groq.com/docs/speech-to-text, read 2026-11-13.
- openai-whisper initial_prompt: github.com/openai/whisper (whisper/transcribe.py); faster-whisper hotwords: github.com/SYSTRAN/faster-whisper (faster_whisper/transcribe.py); read 2026-11-13.
Canonical: https://www.fermionresearch.com/speech/custom-vocabulary-speech-recognition/