Hermes Agent: live dictation does not choose the model that answers
Live dictation joins the 0.22 cycle. Transcription, the answering model and speech playback remain separate services. Check their providers before opening the microphone.
Going deeper
Generated image · SecuFocus
At a glance
Key points
Hermes can show words as you speak and send a voice request to a different model from the one used for typed chat. The intended benefit is a quicker reply, not a conversation that has become local. Check who transcribes the recording, who answers and who turns the answer into speech.
Live dictation needs its own access
Announced in 0.21.6 and detailed in 0.22.0, the optional stt.streaming setting displays partial text as you speak in the CLI, TUI and Desktop dictation. Without it, Hermes waits until recording ends before transcribing the file.
The documentation lists OpenAI, xAI and ElevenLabs for live transcription, each requiring its own API key. Nous-managed transcription still processes recorded files: that access does not open a live transcription session. A subscription or sign-in used for chat does not establish access to the same service for your microphone.
Local Whisper and several other providers continue to process the recording afterwards. If a live connection fails, the documented fallback is ordinary file transcription. SecuFocus has not measured its reliability or delay.
A voice request can use a different answering model
The auxiliary.voice_chat slot selects the model that answers in the chained voice path. It is available under auxiliary models in Desktop and hermes model. The next typed message returns to the session’s model. Voice turns retain access to tools.
By default, this path disables reasoning where supported, or requests the lowest accepted level. That trades analysis time for responsiveness. Keep a model suited to the task when you need substantial reasoning; choosing a faster voice model is not a reason to skip reviewing a sensitive action.
This choice can also change who receives the transcript. A local main model does not make cloud transcription or a cloud voice answering model local.
GPT-Live is a different path with separate billing
Desktop also offers an optional GPT-Live mode that can listen and speak at the same time. The voice layer does not run tools itself. It delegates work to Hermes, using the session’s model and approvals. This mode ignores the auxiliary.voice_chat choice.
The documentation requires an OpenAI API key and describes billing for voice session time, including idle time, in addition to the model calls doing the work. It excludes Nous-managed audio and the CLI, TUI and messaging surfaces from this mode. Check the selected path, not just whether a microphone button exists.
Start with a sentence containing no sensitive information
Dictate a short instruction without passwords or personal data. Check the transcript before requesting an important action. Review the configured transcription, voice-model and speech-synthesis providers, and close the voice conversation when finished.
These are publisher-documented features, not a microphone test by SecuFocus. They establish neither end-to-end confidentiality nor a guaranteed response time.
Check and explore
Sources for this article
Numbers connect each reference to the passages that use it. Dates show when the documentation was consulted.
This article draws on the sources above. The exercises are for you to try on your devices; SecuFocus does not present them as tests carried out by its editorial team. Interfaces and features can change. Method and corrections.
Cite this article
Keep this reference with the article when you save or share it.