hermes gets expressive tts: gemini personas and emotion tags

@iamBarronRoth, who works on gemini at google, added a set of gemini text-to-speech features to nous research's hermes agent. the goal: make the agent's spoken replies sound natural instead of flat and robotic. it's already merged into hermes's main branch

what's new

• directors notes (persona file). write a plain text or markdown file describing how the voice should sound – tone, pacing, character – and hermes attaches it to every tts request. the description shapes the voice but never shows up in the chat reply. set it once, and every spoken response follows the same style

• expressive audio tags. gemini 3.1's tts model supports inline tags that cue emotion and delivery. turn this on and hermes inserts them automatically via a separate rewrite step before generating audio. that step can run on a cheaper, faster model so it doesn't tie up your main chat model. it only activates on gemini 3.1 tts models, and if the rewrite fails, hermes falls back to plain audio instead of erroring

• native telegram voice notes. telegram replies now arrive as ogg/opus voice messages instead of mp3 attachments

what you can build

- an audiobook or long-form narration agent that shifts tone between characters and calm narration, instead of reading every line in the same voice

- a language-tutor bot that actually sounds encouraging, slows down on corrections, and emphasizes the right syllables

- a podcast or news-summary agent that reads with emphasis and pacing instead of a flat tts dump

the details

everything is opt-in and backwards compatible – existing tts setups are unaffected unless you enable it. the persona and audio-tag commits were cherry-picked onto main with roth's authorship preserved