Can I send my Hermes a voice message, and can it answer out loud?
Yes to both, but only answering out loud works out of the box. Hermes Agent turns a voice message into text before the model sees it, and that step needs a transcription provider. A new Vavio computer has none: there is no local Whisper and no transcription key, so a voice message arrives as an audio file Hermes can’t read. If you connect an OpenAI API key, an xAI key or Grok sign-in, or Nous Portal as your model provider, Hermes uses that same account for transcription too, and bills it there. Otherwise the cheapest fix is a Groq key, at $0.04 per hour of audio. Speaking a reply needs no key, because Hermes’s default voices come from Microsoft Edge’s free online service. Either way, the provider you choose bills you directly, separately from your chat model.
This follows Hermes Agent v0.23.3, commit caafe53, checked on 22 November 2026. We reviewed official documentation and source, including how Vavio sets up Hermes, and read prices from OpenAI and Groq that day. We did not send voice messages, so this article makes no claim about accuracy, speed or how the voices sound.
Why can’t my Hermes understand my voice message?
Usually because no transcription provider is ready. Transcription is switched on by default. When a voice message arrives and you haven’t picked a provider, Hermes tries them in this order and uses the first one that works:
- Local Whisper on the computer itself (the
faster-whisperpackage, or awhispercommand). - Groq, if
GROQ_API_KEYis set. - OpenAI, if
VOICE_TOOLS_OPENAI_KEYorOPENAI_API_KEYis set, or through a Nous subscription. - Mistral, xAI, ElevenLabs and DeepInfra, each if its key is set.
If none of these is available, Hermes passes the model a short note instead of your words: “voice message could not be transcribed automatically; the audio is available at …”. The note deliberately says nothing about setup, so Hermes may simply tell you it couldn’t hear the message. It won’t install anything for you while it is guessing.
On a Vavio computer Hermes is installed with its standard extras. Hermes’s own install policy leaves local Whisper out of them, so it downloads only when you choose it, and the Vavio setup adds no transcription key. So until you pick a provider, every voice message ends up as that note. There is one exception. If the model provider you connected is the OpenAI API, xAI (by key or Grok sign-in) or Nous Portal with Tool Gateway access, Hermes finds that account while auto-detecting and transcribes with it, with no voice setting of its own. Those voice messages are then billed to that account. OpenRouter, Anthropic, Google AI Studio, MiniMax and a ChatGPT/Codex sign-in don’t count.
Which transcription provider should I choose, and what does it cost?
| Provider in Hermes | What you need | Who bills you | Price we read on 22 Nov 2026 |
|---|---|---|---|
| Groq | A Groq key | Groq (free tier available) | Hermes’s default whisper-large-v3-turbo: $0.04 per hour of audio, billed in 10-second minimums |
| OpenAI | An OpenAI API key | OpenAI API usage | gpt-4o-mini-transcribe about $0.003 a minute; whisper-1 (Hermes’s default here) $0.006 a minute |
| Local Whisper | Nothing; you choose it in Hermes | Nobody | Free. The default base model is a ~150 MB download on first use and runs on the computer’s processor |
| Nous Subscription | Nous Portal sign-in with Tool Gateway access | Your Nous subscription | OpenAI transcription, billed per use against the subscription |
| Mistral, xAI, ElevenLabs, DeepInfra | That provider’s key | That provider | Not checked |
To put the prices in scale: a 30-second voice note costs about a thirtieth of a cent on Groq, which bills a minimum of ten seconds per message, and about $0.0015 on OpenAI’s gpt-4o-mini-transcribe. A ChatGPT or Codex sign-in doesn’t cover transcription. That sign-in works for Hermes’s image tool, but voice needs an OpenAI API key, which is billed separately from a ChatGPT plan.
Local Whisper costs nothing, but it has to be chosen. Picking Local in Hermes’s settings is what makes Hermes download faster-whisper and the model. Hermes never installs it while auto-detecting. It works without a graphics card, using the computer’s processor. Larger models are more accurate but slower: small is ~500 MB and large-v3 ~3 GB. We didn’t measure speed on a Vavio computer.
Speak another language? Hermes’s transcription setting defaults to English (stt.language: en), because automatic detection misreads short clips. Set your language, or a blank value for automatic detection, under stt.language or in the provider’s own settings.
How do I get Hermes to answer out loud?
Hermes can already speak: its text_to_speech tool is on by default. The default voice provider is Edge: Microsoft Edge’s free online read-aloud voices, with no key or account. Hermes downloads the small edge-tts package the first time it speaks. Ask “read that back to me as a voice message” and Hermes makes an audio file. You can also set it to reply out loud on its own:
- On Telegram and other chat apps, send
/voice onin that chat and Hermes answers your voice messages with voice./voice ttsspeaks every reply, and/voice offgoes back to text. The setting lasts across restarts. To make spoken replies the default for every chat, setvoice.auto_tts: true; it is off on a new install. On Telegram, Hermes converts the audio to a voice bubble, and if conversion fails it sends an ordinary audio file instead. Hermes also shows what it heard, as a 🎙️ message, unless you turnstt.echo_transcriptsoff. - In the Hermes app, turn on Read replies aloud in the message box or under Settings → Voice. The app also has a microphone button and a Voice conversation mode.
If you want different voices, Hermes supports eleven providers, including OpenAI, ElevenLabs, Google Gemini, Mistral and xAI. OpenAI’s tts-1 costs $15 per million characters, so a 1,000-character reply is about $0.015. Its newer gpt-4o-mini-tts, Hermes’s OpenAI default, is billed by tokens: $12 per million audio tokens on 22 Nov. A paid Nous subscription can supply OpenAI voices without a separate OpenAI key, billed per use.
How do I set it up?
- In the Hermes app, open Capabilities and find speech-to-text and text-to-speech under See, hear and create. Choose a provider for each, and paste a key if it needs one.
- From a terminal,
hermes toolsopens the same choices. - Then test it. Send a short voice message and check that the 🎙️ transcript matches what you said. Then ask for a spoken reply.
A key is a secret, so enter it in Hermes’s own settings, not in a chat message. See where to put a secret your Hermes needs.
The microphone in a browser. When you use the Hermes app in a browser, your recording goes to the computer running Hermes, and that computer transcribes it with the provider you chose. So the browser needs no key of its own, but the computer needs a transcription provider. Browsers allow the microphone only over a secure connection. Vavio serves your Hermes app over HTTPS, so your browser shows its usual permission prompt. We haven’t tested the microphone on a Vavio computer.
Hermes’s own talk-and-listen mode, /voice in a terminal, needs a microphone and speakers on the machine running Hermes. On a cloud computer there are none, so use a chat app or the Hermes app instead. Discord voice channels are supported too, through a Discord bot.
Will I see what voice costs me?
Not in Hermes’s own usage figures. hermes insights and the dashboard’s Usage analytics record model tokens. A reply Hermes chooses to speak shows only as a count of text_to_speech calls, with no price. Replies spoken automatically after /voice on or /voice tts don’t appear at all, and neither does transcription of incoming voice messages. Check the provider’s own dashboard instead. With Nous, the Nous Portal dashboard shows usage by tool.
What changes on a Vavio computer?
Vavio doesn’t add a voice provider or pay for transcription. A new computer can speak through Edge but can’t understand voice until you choose a transcription provider, or connect an OpenAI API, xAI or Nous Portal model account that doubles as one. Any paid provider bills your own account. Vavio’s own controls don’t change voice settings either: from Claude, Codex or the vavio CLI you manage the computer and the model, and you choose voice providers in the Hermes app that Vavio opens for you. The computer stays online, so a voice message you send from your phone gets an answer while your laptop is closed. Setting up a chat app such as Telegram happens in Hermes; see Hermes’s voice guide for each platform.
Vavio provides a managed Hermes computer for $14.99 per month, with model and voice usage billed separately by your providers. Compare the current hosting offer, or read what Hermes costs to run each month.
Sources
- Hermes v0.23.3 voice and text-to-speech guide, checked 22 November 2026
- Hermes v0.23.3 voice mode guide, checked 22 November 2026
- How Hermes picks a transcription provider, checked 22 November 2026
- What Hermes tells the model when a voice message can't be transcribed, checked 22 November 2026
- Hermes's install extras, which leave out local Whisper and Edge voices, checked 22 November 2026
- Hermes v0.23.3 Nous Tool Gateway guide, checked 22 November 2026
- edge-tts, the library behind Hermes's default voices, checked 22 November 2026
- OpenAI API prices for transcription and speech, checked 22 November 2026
- Groq speech-to-text prices, checked 22 November 2026
- Vavio hosting offer, checked 22 November 2026
Written by Vavio with AI assistance and checked against pinned official Hermes v0.23.3 documentation and source, against Vavio's own computer setup source, and against provider price lists read on 22 November 2026. This is a source review. We did not send voice messages or judge voice quality.