Freehand

Freehand is a lightweight Windows and macOS client for speech-to-text and text-to-speech services you choose. Dictate into apps, transcribe recordings, or generate spoken audio—with models running on your PC, your network, or a compatible hosted service.

Free forever. Open source. No Freehand subscription or account. Your chosen provider or hosting may have its own costs. You do not need a dedicated local GPU: send inference to another machine to keep this PC's memory and GPU available for your other work, or run your models locally if you prefer.

On Windows and macOS, you can let Freehand install and manage a local NeMo speech runtime, or connect to infrastructure you already run. The recommended local setup uses Nemotron 3.5 streaming for live transcription, with optional MagpieTTS speech generation in the same runtime. Runtime and model downloads are optional and separate from the app; performance depends on your hardware, model, network, and optional cleanup stage.

Download Freehand · Get started · Read the documentation

Choose your workflow

• Voice dictation — use a global shortcut to capture speech and insert the transcript into the original application when the target remains safe. Otherwise, the result stays available to copy.
• Audio file transcription — select a recording, transcribe it, and copy the result explicitly. No microphone or recording shortcut is required.
• Text to speech — enter text, generate spoken audio, and play or explicitly save it. This optional task uses its own speech-generation service; it does not require a transcription connection.

Dictation is the default task, not a setup requirement for the others. Microphone and file transcription have independent connection/model settings. Optional live dictation is a mode of the Voice transcription selection with qualified Nemotron / NeMo-Speech.cpp or Qwen3-ASR and Voxtral / vLLM profiles, showing live results and single-row overlay captions. Cleanup and text-to-speech can each use a separate service and model.

Optional transcript cleanup

Leave transcripts unchanged, use custom instructions with a separate chat model, or select the S1-mini by Superwhisper profile for its English style, structure, and context controls. Use your own inference service or explicitly install the managed llama.cpp runtime and S1-mini model. If cleanup fails or returns empty text, Freehand falls back to the raw transcript.

Configure transcript cleanup

Highlights

• Toggle recording or hold to talk across Windows and macOS applications.
• Use independent OpenAI-compatible endpoints for speech recognition, optional transcript cleanup, and optional speech playback.
• Install an optional local speech runtime on Windows or macOS, browse its supported model catalog, and use Nemotron streaming without configuring a server URL.
• Keep the current transcript available to copy with history disabled.
• Use local voice detection for silence trimming, automatic stop, and pause-aware checkpoints.
• Insert voice transcripts only when the original target remains safe, or copy them explicitly.
• Opt into bounded, memory-only history; configure or disable the native status overlay.

Freehand does not bundle models in its executable. Managed runtime installation is optional; managed NeMo and llama.cpp support Windows and macOS, while managed whisper.cpp remains Windows-only.

Install

Freehand supports Windows 11 x64 or ARM64 with WebView2 and macOS 13 or newer on Apple Silicon or Intel (managed llama.cpp requires macOS 13.3). Use a managed local speech runtime or a reachable compatible service for your task. Only dictation needs a microphone; recording shortcuts are optional. Download the Windows installer/portable executable or matching-architecture macOS ZIP from GitHub Releases. Older releases may not include every architecture; check the assets attached to the selected release. On Windows ARM64, connect a compatible service: Freehand's managed Windows runtime recipes currently support x64 only.

Windows builds are not Authenticode-signed, so Windows may identify their publisher as unknown. Verify manual downloads against the published SHA256SUMS file.

Follow Get started to choose a task, connect its service, and complete your first run.

macOS ZIPs contain an ad-hoc-signed app, not a Developer ID-signed or notarized release. Verify against SHA256SUMS from the same release before opening. Follow the macOS install, permissions, and update guide. Source-build instructions are in CONTRIBUTING.md.

Documentation

• Install and update on Windows
• Connect a speech server
• Set up local speech on Windows or macOS
• Use Freehand
• Privacy and safety
• Troubleshooting

Contributing

Bug reports, interoperability results, documentation fixes, design feedback, and code contributions are welcome. See CONTRIBUTING.md for development setup and pull-request guidance.

Please report vulnerabilities through GitHub private vulnerability reporting, not a public issue. Freehand is available under the MIT License. See Freehand project identity for official sources and guidance on the project name and logos.