nodecode-gmi-speech

GMI Cloud speech: say text into an audio file through GMI's request queue, list the voices a model declares, and the model ids the key may call. A Nodecode rewrite of zkasuran/hermes-gmi-media by zkasuran.

GMI's queue does not answer with audio: it accepts a job, runs it, and names a URL when it is done. The cell does that round trip for one utterance, checks every input against the model's own schema before submitting anything, and answers where the recording landed.

Get it

Tell the agent "add the gmi-speech cell", or run

nodecode add nc://gmi-speech

Set it up

Make a key at https://console.gmicloud.ai, then either paste it into the cell's api_key in /settings, or export it:

export GMI_API_KEY=...

The defaults are the plugin's: minimax-tts-speech-2.6-turbo, then inworld-tts-1.5-mini when no model is chosen, an mp3 unless the model offers something else, and the recordings under voice/ in the profile's home. What is worth setting lives in the gmi-speech section, which /settings writes:

{
  "gmi-speech": {
    "api_key": "",
    "api_key_env": "GMI_API_KEY",
    "model": "minimax-tts-speech-2.6-turbo",
    "voice": "English_expressive_narrator",
    "speed": "",
    "format": "mp3",
    "directory": "voice/",
    "api_base_url": "https://console.gmicloud.ai/api/v1/ie/requestqueue/apikey",
    "poll_timeout_seconds": 300
  }
}

Setting	Default	What it does
`api_key`	none	the GMI Cloud key, written in the section
`api_key_env`	`GMI_API_KEY`	the variable read when `api_key` is empty
`model`	none	the speech model every call uses; empty, the first of the plugin's two GMI accepts
`voice`	none	the voice every call speaks in unless it names one; empty, the model's own default
`speed`	none	the pace asked for; the model clamps it to its own range
`format`	`mp3`	mp3, wav, flac or ogg; a model that cannot write it falls back to one it can
`directory`	`voice/`	where the recordings are kept, under the home
`api_base_url`	GMI's queue	change it only for a stand-in
`poll_timeout_seconds`	`300`	how long one job is followed; 30 to 1800

Use it

gmi_speech_say — {text, voice?, speed?, format?, model?}. Reads the model's schema, builds the payload its parameters name, submits it, follows the job and downloads the recording to a fresh file whose extension is the container the model declared. The answer names that path: give it to the operator.
gmi_speech_voices — {model?, query?}. The voices the model's live schema declares, with the section's own marked; when a model declares none, the two names the plugin offered. One schema read, nothing generated.
gmi_speech_models — {query?}. The ids GMI's queue lists for this key, with the queue's own uuid and probe ids dropped. One request.
/gmi-speech — the settings; /gmi-speech voices [model], /gmi-speech models [words], /gmi-speech say <text>.
(help :gmi-speech) — the model's page.

Ask it "read this sentence aloud with GMI" and it calls gmi_speech_say; ask which voices the model has and it calls gmi_speech_voices first.

What is in it

File	What it is
`package.lisp`	the package, GMI's queue and parameter names as data, the shared helpers
`cell.lisp`	the section, the three tools, `/gmi-speech`, `(help :gmi-speech)`, the schema/payload and job logic
`test/cell-test.lisp`	lifecycle, every tool and the command, every refusal, on a stubbed `nlk:http`
`config.example.jsonc`	the section, commented

Why three tools, not a provider

In Hermes the plugin registered itself as the harness's text-to-speech provider, so every spoken reply went through it. 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. GMI's queue is not that wire -- a job is submitted and polled, and the audio arrives as a URL -- so what a user reaches for here is a recording on disk, the voices a model declares, and the ids there are: three tools and a command.

Test

From the cells repository's root, offline:

.github/test-cell.sh tools/nodecode-gmi-speech

nc://gmi-speech