stt-vocab

A Hermes plugin that tells the speech-to-text model which names to expect. Whisper-family models mis-hear the words that matter most in a voice message: people, products, projects, acronyms. They spell them right far more often when the request carries a short prompt that contains them. Hermes can send one static prompt (stt.prompt); this plugin builds it for you, per request, from:

• the terms you list in its settings, globally or for one surface (gateway voice messages, voice mode), and
• names it finds in this profile's memory files (memories/USER.md and memories/MEMORY.md), the notes Hermes already keeps about you. Off for cloud STT unless you allow it (see Privacy).

stt.prompt (yours, kept as is)      Hermes, Nous Research.
+ stt-vocab                         PostgreSQL, Kubernetes, Project Atlas, Northwind, Ayşe Demir, Bartholomew.

It uses Hermes's pre_transcription hook, so it works with every built-in backend that accepts a prompt: local faster-whisper, OpenAI, Groq, Mistral and DeepInfra. Backends that do not take a prompt (xAI, ElevenLabs, command providers) transcribe as before.

Install

hermes plugins install stt-vocab --enable
hermes gateway restart   # if you use voice messages through the gateway

Settings

All optional, under plugins.entries.stt-vocab.settings. They are read on every transcription, so a change applies without a restart.

Setting	Default	Meaning
terms	[]	Terms for every transcription.
source_terms	{}	Extra terms for one caller surface: gateway (voice messages), voice_mode, ... These come first.
from_memory	true	Add names found in memories/USER.md and memories/MEMORY.md.
allow_remote	false	Also send memory-derived names to cloud STT providers.
max_chars	896	Longest prompt the plugin builds (1–4000). 896 is what Hermes keeps for Whisper-family backends, so it never has to cut the prompt.

plugins:
  entries:
    stt-vocab:
      settings:
        terms: ["Teknium", "Kanban"]
        source_terms:
          gateway: ["Ayşe", "Northwind"]

Order matters: Whisper conditions on the end of the prompt, so the plugin writes the least important terms first and the most important last: memory names, then terms, then source_terms. Your stt.prompt stays at the front untouched, and terms it already contains are not repeated. When everything does not fit in max_chars, memory names are dropped first.

A setting the plugin cannot use (for example terms: "Atlas" instead of a list) is logged once as a warning and the transcription goes out unchanged.

How names are found in memory

Memory entries are free text, so the plugin keeps the words that look like names:

• words with a capital letter after the first one (gRPC, PostgreSQL, AWS),
• words that mix letters and digits (S3, GPT-5),
• capitalised words in the middle of a sentence (Northwind),
• runs of these as one name (Ayşe Demir, Project Atlas, GitHub Actions).

A capitalised word that only opens a sentence is not trusted on its own. Common sentence openers (Prefers, Uses, Then, ...), words that also appear in lowercase in your notes, weekday and month names, URLs, paths, e-mail addresses and long letter-and-digit strings that look like keys are never sent. More frequent names win when space is short. The rules are tuned for notes written in English; in languages that capitalise every noun (German) more ordinary words will pass.

Privacy

The prompt is uploaded to the STT provider together with the audio. Hermes's own documentation asks that nothing session-derived goes into it for a hosted API, so by default:

• your terms and source_terms go to every provider, like stt.prompt does;
• memory-derived names go only to a backend running on this machine: local (faster-whisper) or local_command. For any other provider, including type: command providers whose destination the plugin cannot know, the memory files are not even read.

Set allow_remote: true to send memory-derived names to cloud providers too.

Security and footprint

• register() makes a single call, ctx.register_hook("pre_transcription", ...). No tools, middleware, commands or slash commands.
• No network access and no subprocesses. Hermes, not the plugin, sends the request.
• Reads only its own settings and, when memory names are on and the backend is local (or allow_remote is set), the active profile's memories/USER.md and memories/MEMORY.md, at most 256 KB each. Writes nothing.
• Returns at most a new prompt; never changes model, language or the audio file. Hermes runs the hook fail-open: if it raised, the request would go out unchanged.

Compatibility

Needs Hermes 0.21.0 or later (pre_transcription and the stt.prompt threading shipped there). Tested against v2026.8.31 (0.21.0) and main.

Development

python -m pytest tests/test_terms.py tests/test_plugin.py   # no Hermes needed
PYTHONPATH=/path/to/hermes-agent /path/to/hermes-agent/.venv/bin/python -m pytest tests

tests/test_hermes.py installs the plugin into a fresh HERMES_HOME, enables and configures it with the Hermes CLI, writes memory files where Hermes keeps them, and runs Hermes's own transcribe_audio against a local fake OpenAI-compatible server that records the prompt field of each request.

License

MIT