# Whisperstream

> Private push-to-talk dictation app for Windows and Mac that transcribes speech locally on your computer.

Whisperstream is a free, open-source dictation app from Lanreal Technologies Inc. It turns speech into text in the focused field of other apps, with core transcription running on the user's own Windows PC or Apple silicon Mac. The Mac version is currently in beta.

## What It Is

Whisperstream is a push-to-talk dictation tool. You hold a hotkey (Right Shift by default), speak, and release; the transcribed text is delivered to the active field by clipboard paste or optional simulated keystrokes. Transcription uses open-weight speech models (NVIDIA Parakeet, Qwen3 ASR or Cohere Transcribe) and supports 39 languages. After the one-time model download, core dictation works without an internet connection.

## Optional AI Cleanup and Customization

Cleanup is off by default. When enabled, it removes filler words, fixes grammar and punctuation, and applies spoken self-corrections. It can run on Stream-1 Mini, an on-device model fine-tuned from Qwen3 0.6B, on a larger Customizable local model, or through a provider the user configures (Gemini, OpenAI, Anthropic, Ollama or compatible services). Per-app profiles apply custom prompts automatically by application or website, and a custom dictionary boosts, replaces or filters terms, with CSV import.

## Privacy and Workflow Features

The README states audio is never uploaded and that the app has no telemetry or analytics. Text leaves the computer only if the user enables cloud cleanup. Transcript history is encrypted on the device and searchable, with audio replay. Lockdown Mode limits cleanup to on-device models, turns off cloud cleanup and protects history. Agent Speak connects to Claude Code and Codex so replies can be read aloud and answered by voice. Audio files can also be imported and transcribed locally.

## Setup Path

The official installer supports Windows 10 and 11 (x64) and runs on CPU without a GPU. The Mac build needs Apple silicon and macOS 14 or later; Intel Macs are not supported. First-run setup covers language choice, shortcut, model download and an optional enhancement step. The project is GPL-3.0 licensed and can be built from source.

## Features
- On-device speech transcription
- Push-to-talk and toggle hotkeys
- Text delivery into any app via paste or keystrokes
- 39 languages
- Optional AI cleanup with Stream-1 Mini
- Per-app profiles with custom prompts
- Custom dictionary with CSV import
- Agent Speak for Claude Code and Codex
- Lockdown Mode
- Encrypted transcript history with audio replay
- Audio file import
- System-volume ducking
- Offline operation after model download
- Runs on CPU without a GPU

## Integrations
Claude Code, Codex, Ollama, OpenAI, Anthropic, Gemini, NVIDIA Parakeet, Qwen3 ASR, Cohere Transcribe, sherpa-onnx, llama.cpp

## Platforms
WINDOWS, MACOS, LINUX, API, VSC_EXTENSION

## Pricing
Open Source

## Links
- Website: https://whisperstream.io
- Documentation: https://whisperstream.io/docs
- Repository: https://github.com/lanreal-io/whisperstream
- EveryDev.ai: https://www.everydev.ai/tools/whisperstream