Hermes Agent skill: wire the microphone to an OpenAI-compatible cloud STT endpoint (Polza.ai, Groq, DeepInfra, any aggregator) so dictation is fast and accurate in Russian, instead of a slow local whisper.
Covers the profile trap (the desktop app often runs a named profile, not default), the
provider-slot choice that survives without a gateway restart, and measured latency/quality
for six backends.
It also ships a working brand-name normalizer (scripts/asr_normalize.py +
templates/variants.txt): a type: command STT provider that calls the endpoint itself and
then rewrites the transcript, because stt.prompt is only a soft decoder bias and never
reliably turns a Russian-sounding brand («юджайл», «Фьюджайли», «Fujile», «FiuJail») into
YouGile. Matching is on sound, not spelling: the explicit variant list handles the known
mishearings, a consonant-skeleton pass catches the spellings the model invents next.
hermes skills install degin/hermes-voice-stt-cloud/skills/autonomous-ai-agents/hermes-voice-stt-cloud
Preview without installing:
hermes skills inspect degin/hermes-voice-stt-cloud/skills/autonomous-ai-agents/hermes-voice-stt-cloud
The skill loads in a new session (the skill loader is cached per session).
- Hermes Agent with the
openaipython package (ships with it) andffmpegon PATH - An API key for any OpenAI-compatible
/audio/transcriptionsendpoint - The key never lives in the skill: it goes into the profile's
.env
Text-to-speech and the Hey Hermes wake word are separate Hermes sections.