HoldSpeak is a local, cross-platform AI copilot for macOS and Linux, offering advanced dictation and meeting transcription. It targets power users and engineers prioritizing data privacy and workflow automation, enabling text input anywhere and generating structured meeting artifacts without cloud reliance.

How It Works

The system processes voice locally using Whisper for transcription. Users can then leverage their own LLM (GGUF, MLX, OpenAI-compatible) for context-aware rewriting of dictation or for extracting structured data like decisions and action items from meeting recordings/transcripts. A key design choice is the "nothing leaves your machine" principle, ensuring all sensitive data remains local. The dictation mode features a learning loop where corrections improve future transcriptions, while the meeting mode offers an approval-gated workflow for generated action items.

Quick Start & Requirements

• Install: pip install holdspeak or uv pip install holdspeak. Install extras like pip install 'holdspeak[meeting]' or pip install 'holdspeak[dictation-llama]'.
• Prerequisites: Python, a local LLM (see docs/MODELS.md), microphone access.
• Setup: Run holdspeak doctor to check permissions and backends.
• Docs: Official quick-start guides and documentation are available at: Documentation index, Architecture, Getting Started, and Models.

Highlighted Details

• End-to-End Local AI: Transcription (Whisper) and LLM inference run entirely on the user's machine.
• Adaptive Dictation: Features a dictation journal and correction memory that learns from user feedback, allowing replay to observe pipeline improvements.
• Meeting Intelligence: Automatically generates typed artifacts (decisions, action items, ADRs) from meeting transcripts via LLM plugins, with an auditable approval flow for actions.
• Flexible LLM Integration: Supports various local LLM formats (GGUF, MLX) and OpenAI-compatible endpoints.

Maintenance & Community

Contributions are welcomed, with guidelines provided in CONTRIBUTING.md. Recent changes are tracked in CHANGELOG.md. No specific community channels (e.g., Discord, Slack) or major contributor/sponsor information is detailed in the README.

Licensing & Compatibility

Licensed under the Apache License 2.0. This permissive license allows for commercial use and integration into closed-source projects without copyleft restrictions.

Limitations & Caveats

The project is in early 0.x status, meaning APIs and configurations may change. It lacks a Windows build. Wayland support for global hotkeys and cross-app typing is best-effort. Users must provide and configure their own LLM for advanced features. Setup is more involved than typical single-purpose applications.