nodecode-speech
AllModels Speech: one key for text-to-speech, speech-to-text and voice search -- the agent says a sentence into an audio file, reads an audio file back as text, and finds the voice before it speaks. A Nodecode rewrite of allmodels-io/hermes-speech by allmodels-io.
Get it
Tell the agent "add the speech cell", or run
nodecode add nc://speech
Set it up
Create an AllModels API key at https://allmodels.io, then either paste it into the cell's api_key in /settings, or export it:
export ALLMODELS_API_KEY=...
Pick a voice with allmodels_speech_voices (or /speech voices British female narrator) and keep its id in tts_voice and the provider that serves it in tts_voice_provider -- the two travel together. A call can name a voice of its own instead.
The section, as ~/.nodecode/config.jsonc carries it:
| Setting | Default | What it does |
|---|---|---|
api_key |
(unset) | the key, written inline as a secret |
api_key_env |
ALLMODELS_API_KEY |
the variable read when api_key is empty |
tts_model |
fish/s2-1-pro |
the text-to-speech model every call uses unless it names one |
tts_voice |
(unset) | the voice id every call speaks in, when it names none |
tts_voice_provider |
(unset) | the provider that serves tts_voice |
tts_speed |
(unset) | the pace, 0.25 to 4.0, as a number |
tts_format |
mp3 |
mp3, opus, aac, flac, wav or pcm |
stt_model |
soniox/stt-async-v5 |
the speech-to-text model every transcription uses unless it names one |
stt_language |
(unset) | the language recordings are mostly in; unset lets the model detect it |
stt_prompt |
(unset) | names and jargon the recordings are likely to contain |
api_base_url |
https://api.allmodels.io |
AllModels' API root; the audio endpoints are its /oai path |
directory |
speech/ |
where the synthesized audio is kept, under the home |
Without a key the tools still answer, with the sentence that says which setting to fill, and nothing is billed. Voice search needs no key at all.
Use it
Ask for speech: "say this out loud and give me the file", "read this voice note back to me", "find a warm British voice". The model calls:
| Tool | What it does |
|---|---|
allmodels_speak(text, voice?, voice_provider?, model?, speed?, format?) |
one utterance, saved as an audio file in the speech folder; the answer names the path |
allmodels_transcribe(path, model?, language?, prompt?) |
one audio file on this machine, as the words it says |
allmodels_speech_voices(query, model?, provider?, language?, limit?) |
the public AllModels voice catalogue, searched in words; no key, nothing billed |
The /speech command answers what the cell is set to (/speech), the account's credit (/speech balance), and the same voice search the tool runs (/speech voices <words>). (help :speech) is the manual topic the model reads.
The audio endpoints are AllModels' OpenAI-compatible surface -- POST /oai/v1/audio/speech and POST /oai/v1/audio/transcriptions -- so the same key that serves speech serves the models next door. Every call that reaches AllModels is billed to the account the key belongs to.
Credit
This cell is a rewrite of allmodels-io's hermes-speech; see NOTICE. MIT, see LICENSE.
