nodecode-kokoro-tts

Kokoro: a local voice, through a Kokoro server on loopback — text in, a WAV file out, one voice per sentence's language. A Nodecode rewrite of brunosouzas/hermes-kokoro-tts by brunosouzas.

Nothing leaves the machine: the cell speaks to the same Kokoro server the plugin ran, on 127.0.0.1, and answers where the recording is. With a voice map one answer can be read in several languages, each sentence in the voice of its own.

Get it

Tell the agent "add the kokoro-tts cell", or run nodecode add nc://kokoro-tts.

Set it up

The cell needs a Kokoro server, not a key. The plugin's own (hermes kokoro install, then hermes kokoro restart) listens on http://127.0.0.1:8770 and serves /v1/audio/speech and /v1/voices, which is what the cell dials. Any other Kokoro server speaking that OpenAI-shaped /v1/audio/speech answers too: point url at it.

The defaults are the plugin's, so with that server running there is nothing to set. What is worth setting lives in the kokoro section, which /settings writes:

{
  "kokoro": {
    "url": "http://127.0.0.1:8770",   // the server's root
    "voice": "af_heart",              // the default voice
    "voices": ["pt=pf_dora"],         // a language of its own, LANGUAGE=VOICE
    "speed": "1.0",                    // 0.25 to 4.0
    "output_dir": "~/voices"           // where recordings go; unset, the cache
  }
}

Setting	Default	What it does
`url`	`http://127.0.0.1:8770`	the Kokoro server's root
`voice`	`af_heart`	the voice everything speaks with unless something names another
`voices`	none	one voice per language, `LANGUAGE=VOICE`: `pt`, `en` and `it` are detected, and `it` only when it has a voice
`speed`	`1.0`	how fast the voice speaks, 0.25 to 4.0
`language`	none	single-voice mode: read as this Kokoro language code (`p`, `pt-br`) instead of the voice's own
`english_words`	none	English words a non-English voice should say in English
`english_accent`	`a`	the accent those loanwords are read with: `a` American, `b` British
`output_dir`	the cache	where recordings are written
`timeout_seconds`	`120`	one request's ceiling

A voice is a Kokoro voice id — a language letter, a gender letter and a name (af_heart, bf_emma, pf_dora, if_sara). Anything else is refused at start, so a voice can never name a file.

Use it

kokoro_speak — {text, voice?, path?}. Says the text, one request per language run, joins the recordings into one WAV and answers where it is, how long it plays and which voices spoke. path is where the file goes; unset, a fresh file under the cell's cache.
kokoro_voices — {language?}. What the server holds, each voice with the language it reads and which one the section speaks with.
/kokoro TEXT — the same, for the operator. /kokoro with no text says where the server is, what will speak, and the three speech settings that give a room's spoken replies the same voice.
(help :kokoro) — the model's page.

A room's spoken replies already speak through an OpenAI-compatible /v1/audio/speech, so the same server can do both: speech.base_url http://127.0.0.1:8770/v1, speech.model kokoro, speech.voice your pick. The cell and the room then share one voice, and only the cell does the per-sentence language routing.

What is in it

File	What it is
`cell.lisp`	the section, the two tools, `/kokoro`, `(help :kokoro)`, the detector and the WAV joiner
`test/cell-test.lisp`	lifecycle, both tools, the command and the refusals, every request a stubbed `nlk:http`
`config.example.jsonc`	the section, commented

Why two tools, not a provider

In Hermes the plugin registered itself as the harness's text-to-speech provider, and its own server did the speaking. Nodecode's provider seam is models, not voices: speech in a room comes from the channel kit's speech section, which takes an OpenAI-compatible base_url, model and voice — so a running Kokoro server is already one configuration line away, and what was missing is a way to ask for a recording and to see which voices there are. Those are the two tools. The per-sentence language routing, the voice id rule and the server's request shape are the original's.

Test

From the cells repository's root, offline:

.github/test-cell.sh tools/nodecode-kokoro-tts

nc://kokoro-tts