chuchote 0.1.1

Local-first, hands-free voice assistant for Ollama.

Chuchote

Local-first, hands-free voice assistant for Ollama. Talk to a local LLM and get a spoken answer — wake word → speech-to-text → local reasoning → text-to-speech, running entirely on your machine. No cloud, ever.

Status: Phase 5 — the CLI MVP is complete. Wake word → speech capture (Silero VAD end-of-turn) → faster-whisper → Ollama → Piper → speaker, with memory across restarts, barge-in to interrupt a reply, overlapped synthesis, a config file, and an install script. Push-to-talk remains a fallback (--ptt). See CLAUDE.md for what's next.

Pipeline

wake word → capture (Silero VAD end-of-turn) → faster-whisper (STT)
          → Ollama (reasoning) → Piper (TTS) → speakers  ↺

The wake word (openWakeWord) and VAD (Silero) both run locally on onnxruntime.

Requirements

• Python 3.10+
• Ollama running locally with a model pulled (e.g. ollama pull llama3.2)
• A Piper voice (.onnx + .onnx.json) — drop it in ./voices/
• A working microphone and speakers

Install

python -m venv .venv
# Windows:  .venv\Scripts\activate
# macOS/Linux:  source .venv/bin/activate
pip install -e .

Or use the install script, which also downloads a default Piper voice and writes a config file:

./scripts/install.ps1     # Windows (PowerShell)
./scripts/install.sh      # macOS/Linux

Run

Make sure Ollama is running (ollama serve) and a voice model is in ./voices/:

chuchote start
# or override defaults:
chuchote start --model llama3.2 --wake-word hey_jarvis --voice voices/en_US-lessac-medium.onnx

By default Chuchote is always listening for the wake word (hey_jarvis). Say it, then speak your request — Silero VAD detects when you've stopped and the reply is transcribed, thought out, and spoken back (starting as soon as the first sentence is ready). Press Ctrl+C to quit.

The wake-word models download automatically on first run. Other built-in words include alexa, hey_mycroft, and hey_rhasspy (--wake-word); raise --wake-threshold if you get false triggers. A short tone confirms the wake word — silence it with --no-chime.

Barge-in. While Chuchote is speaking, say the wake word again to cut it off and start a new turn (--barge-in wake, the default — echo-robust, works with open speakers). With headphones you can use --barge-in vad so any speech interrupts; --barge-in off lets every reply finish. In push-to-talk mode, pressing the PTT key cuts off the reply (then hold it to speak again).

Custom wake words

--wake-word accepts a file path as well as a built-in name, so you can use any openWakeWord-compatible model:

chuchote start --wake-word path/to/hey_chuchote.onnx

To create one for your own phrase, use openWakeWord's automatic training notebook — see Training New Models in their README. It generates synthetic speech for your phrase and trains a model in roughly an hour, no voice recordings needed; download the resulting .onnx and point --wake-word (or wake_model in your config) at it.

Note: the training notebook runs on Google Colab (a free cloud notebook) — that's one-time, dev-machine tooling, like downloading a Piper voice. The resulting model runs fully locally; nothing about the assistant touches the network.

Push-to-talk fallback

Prefer holding a key? Skip the wake word entirely:

chuchote start --ptt            # hold space to talk, release to send
chuchote start --ptt --ptt-key ctrl

Checking your setup

Not sure everything's wired up? Run the preflight check:

chuchote doctor

It verifies Ollama is reachable and the model is pulled, a Piper voice is present, your mic and speakers are detected, and (in wake mode) the wake-word deps are installed — reporting each as [ ok ] / [fail] and exiting non-zero if anything's wrong.

Languages

Chuchote isn't English-only — whisper understands ~99 languages and Piper has voices for dozens. To run it in another language, three things need to line up:

1. Recognition — set language and use a multilingual whisper model (the plain names, not the .en ones — base on low-RAM machines, small for better accuracy):

   language = "fr"
   whisper_model = "base"
   

2. Speech — download a Piper voice for that language from VOICES.md into your voices dir (or point piper_voice at it). Piper voices are one language each.
3. Reasoning — pick an Ollama model that's good in your language (most modern ones are multilingual; e.g. qwen2.5 is strong for Chinese). Chuchote already asks the model to reply in whatever language you speak.

Then chuchote doctor will confirm the pieces match. Common starting points:

Language	language	Example Piper voice
French	fr	fr_FR-siwis-medium
German	de	de_DE-thorsten-medium
Spanish	es	es_ES-davefx-medium
Italian	it	it_IT-paola-medium
Portuguese (BR)	pt	pt_BR-faber-medium
Dutch	nl	nl_NL-mls-medium
Chinese	zh	zh_CN-huayan-medium
Russian	ru	ru_RU-dmitri-medium

Adding any other language: find its voice on VOICES.md, set language to whisper's code for it, keep a multilingual whisper_model, and you're set. The wake word stays an English phrase (openWakeWord's built-ins are English) — or use --ptt.

Memory

Chuchote remembers the conversation. Each exchange is stored in a SQLite database (memory.db in your per-user data dir) and the most recent turns are fed back into the model's context every turn, so it stays coherent across turns and restarts.

chuchote start --forget   # start a session with a clean slate
chuchote forget           # erase all saved memory

Tune how much history is injected via history_messages in chuchote/config.py.

Configuration

Persist your settings in a config file instead of passing flags every time:

chuchote init              # writes a commented config.toml to your config dir

Edit the file (its path is printed by init; typically %APPDATA%\chuchote\ on Windows or ~/.config/chuchote/ elsewhere), uncomment what you want to change, and it's picked up on the next chuchote start. Point at a specific file with --config PATH. Precedence is defaults < config file < flags.

Flags (over sensible defaults; see chuchote/config.py for the full list and VAD tunables):

Flag	Default	Meaning
--model	llama3.2	Ollama model
--whisper-model	base.en	faster-whisper model (small.en = more accurate, needs ~1 GB free RAM)
--language	auto	recognition language (en, fr, de, zh, …); needs a multilingual model
--voice	first .onnx in ./voices	Piper voice model
--wake-word	hey_jarvis	wake word model (alexa, hey_mycroft, hey_rhasspy)
--wake-threshold	0.5	wake sensitivity 0..1 (higher = fewer false triggers)
--no-chime	off	disable the wake-word acknowledgement tone
--barge-in	wake	interrupt a reply: wake / vad (headphones) / off
--ptt	off	use push-to-talk instead of the wake word
--ptt-key	space	push-to-talk key to hold (with --ptt)
--forget	off	clear conversation memory before starting
--no-banner	off	don't print the startup banner
--config	per-user config dir	path to a config file

Development

Run the test suite (covers the pure logic — sentence chunking, memory, config precedence, banner styling — with no audio/model deps needed):

pip install -e ".[dev]"
pytest

Tests also run in CI (GitHub Actions) on Linux and Windows, Python 3.10–3.13.

Releasing

Publishing a GitHub release triggers the publish.yml workflow, which builds and uploads to PyPI via trusted publishing — no API token. One-time setup on pypi.org: add a trusted publisher for this repo (owner Cjayy77, repo chuchote, workflow publish.yml, environment pypi), and create a matching pypi environment in the repo's GitHub settings.

License

MIT.