saymo 0.9.0

Local AI voice assistant for macOS — speaks in your cloned voice on live calls, fully on-device

Saymo — Local AI Voice Assistant

Fully local AI voice assistant for macOS. Speaks into any live call in your cloned voice — no cloud APIs required.

Saymo composes short, natural speech from optional data sources (tracker, notes, text files), synthesizes it with voice cloning, and routes audio into the active call through a virtual microphone. Everything — language model, speech-to-text, text-to-speech — runs on-device.

• Local: Ollama + faster-whisper + Coqui XTTS v2 (or Piper / macOS say as fallback).
• Voice cloning: 5-minute sample → your voice, fine-tuning optional.
• Routing: BlackHole virtual mic → any browser-based call app.
• Call automation: Chrome-driven mute/unmute for 8 providers (Glip, Zoom, Google Meet, MS Teams, Telegram, Yandex Telemost, VK Teams, MTS Link).
• Listening mode: auto-detects when your name is called, answers questions from provided context.
• User-configurable prompts and vocabulary — no source edits required.

Project status: early public alpha. Expect rough edges. Contributions welcome.

---

Requirements

• macOS with Apple Silicon (M1/M2/M3/M4), arm64 terminal, not Rosetta
• Python 3.11+
• Homebrew
• Google Chrome
• ~10 GB free disk space

Quick install

git clone https://github.com/mshegolev/saymo && cd saymo
cp config.example.yaml config.yaml   # fill in your details
./install.sh

The installer handles brew deps, Python packages (via uv or pip), an Ollama check, a Piper voice model, and Chrome permissions.

First-time setup

saymo setup                        # Interactive wizard: name, devices, profiles
saymo record-voice -d 300          # Record a 5-minute voice sample
saymo test-devices                 # Verify audio devices
saymo test-tts "Привет, это тест"  # Check that TTS works

One-time audio routing

┌─────────────────────────────────────────────────────────────┐
│                   Audio MIDI Setup                          │
│  Create "Multi-Output Device":                              │
│    ✓ Your headphones   (master, no drift correction)        │
│    ✓ BlackHole 16ch    (drift correction ON)                │
│                                                             │
│  In your call app:                                          │
│    Microphone → BlackHole 2ch                               │
│    Speakers   → Multi-Output Device                         │
└─────────────────────────────────────────────────────────────┘

Daily usage

# Before the call: prepare text + cached audio
saymo prepare -p personal
saymo prepare-responses         # pre-synthesize the Q&A library for live mode
saymo review                    # optional: check generated audio

# During the call
saymo speak -p personal         # manual trigger, instant playback
saymo auto -p personal          # listen for your name, speak when called
saymo auto -p personal --mic    # same, but from laptop mic (for testing)

# Extras
saymo dashboard                 # interactive TUI

Call providers

saymo auto works with all Chrome-based call apps — the provider is picked by meetings.<profile>.provider in config:

provider:	Service
glip (default)	RingCentral Glip
zoom	Zoom
google_meet	Google Meet
ms_teams	Microsoft Teams
telegram	Telegram calls (web)
telemost	Yandex Telemost
vk_teams	VK Teams
mts_link	MTS Link

Run saymo list-plugins to see everything available in your install.

Live Q&A mode

When your name is called and the surrounding transcript looks like a question, auto consults a pre-synthesised response library and plays the best-matching cached variant — no network hop, no synthesis lag. Populate the library once with saymo prepare-responses. Built-in intents cover status (как дела), blockers, ETA, testing stage, review. Extend with your own wording via config.responses.library.

On cache miss, you can opt into a live fallback: Ollama composes an answer from your standup summary + JIRA context, the TTS engine synthesizes it, and Saymo plays it back. This adds a few seconds of latency but covers any question. Enable it in config:

responses:
  live_fallback: true

Without live_fallback (default), a cache miss falls back to the generic standup audio — quiet, reliable, no LLM dependency.

Configurable prompts

All LLM prompts are templates loaded from config.yaml → prompts.* at runtime, with sensible generic defaults in source. To customize voice/tone:

prompts:
  standup_ru: |
    Ты — помощник для ежедневных встреч. Составь отчёт на русском...
    {yesterday_notes}
    {today_notes}
  qa_system_ru: |
    Ты — {user_name}, {user_role}. Отвечай кратко, 1-3 предложения...

See config.example.yaml for all available keys and the default set.

Project-specific vocabulary

Adding your own abbreviations or fuzzy name expansions to the TTS normalizer is done through config, not source:

vocabulary:
  abbreviations:
    MYAPI: "май-эй-пи-ай"
    K8S: "кубернетес"
  fuzzy_expansions:
    Alex: ["Alex", "Алекс", "Саша", "Саня"]

Architecture

┌───────────────┐   ┌──────────────┐   ┌────────────────┐   ┌──────────────┐
│ Source plugin │──▶│ LLM composer │──▶│ Text normalizer│──▶│  TTS engine  │
│  (optional)   │   │   (Ollama)   │   │   (abbrevs,    │   │  (XTTS clone │
│               │   │              │   │    numbers)    │   │  / Piper)    │
└───────────────┘   └──────────────┘   └────────────────┘   └──────┬───────┘
                                                                   │
┌──────────────┐   ┌──────────────┐   ┌────────────────┐           │
│Call provider │◀──│ Auto trigger │◀──│  STT (Whisper) │       Audio bytes
│(mute/unmute) │   │(name detect) │   │ (capture call) │           │
└──────┬───────┘   └──────────────┘   └────────────────┘           │
       │                                                           │
       ▼                                                           ▼
  BlackHole 2ch ─────────────────────────────────────────── Audio output + monitor
  (virtual mic)

Details in docs/PRD.md and ADRs under docs/adr/.

Security & privacy

• Everything runs on-device by default. Cloud TTS / STT providers are optional and disabled in the example config.
• Voice samples and secrets are listed in .gitignore — they never leave your machine.
• Prompts, vocabulary, trigger phrases are all in your config file — source stays generic.

License

MIT — see LICENSE.

Acknowledgements

• Coqui TTS for XTTS v2.
• Ollama for local LLM hosting.
• faster-whisper for transcription.
• BlackHole for virtual audio routing.