Eqho — free, offline voice-to-text dictation

Speak anywhere Windows lets you type. Press a hotkey, talk, and your words appear in whatever app is focused — Word, your browser, Slack, code editors, anything. Powered by OpenAI's Whisper (via faster-whisper and whisper.cpp), running 100% on your machine — accelerated on any modern GPU: NVIDIA, AMD, or Intel.

• 🔒 Private by design — no cloud, no account, no API key. Your voice never leaves your computer.
• ⚡ Fast on any GPU — NVIDIA (CUDA), AMD and Intel (Vulkan), or CPU. Eqho auto-detects your hardware and picks the fastest local engine; real-time preview as you speak.
• 🆓 Free and open source — no subscription, no word limits, ever.
• 🌍 13 languages — English-optimized by default, multilingual models one click away.

Install

Windows (recommended): grab Eqho-Setup-<version>.exe from Releases → double-click → done. Prefer no installer? Use the portable .zip.

Linux / macOS: download the .tar.gz / .dmg from Releases (core dictation supported since v0.6.0 — see platform notes).

From source:

git clone https://github.com/danielmevit/eqho.git
cd eqho
python -m venv venv && venv\Scripts\activate   # Windows
pip install -r requirements.txt
python run.py

First run downloads the default speech model (distil-large-v3, ~1.5 GB) — after that, everything is offline.

How it works

1. Click into any app where you want text.
2. Press Alt+Q — a floating bar shows "Listening…".
3. Speak naturally. Watch the live transcription preview.
4. Press Alt+Q again — your words are typed into the app.

Toggle or hold-to-talk, your choice. The hotkey is customizable.

Features

• Dictation into any application via clipboard paste or simulated typing
• Live transcription overlay with adjustable position and opacity
• Settings dashboard — dark, light, and system themes
• Transcript history — searchable, exportable, stored locally, one-click clear
• Custom vocabulary — teach it your names and jargon
• Voice commands — "new line", "period", "delete that" (opt-in)
• Text replacements — auto-correct words the model keeps missing
• Volume ducking — quiets your speakers while you dictate
• Model picker — from tiny (150 MB, fastest) to large-v3 (3.1 GB, most accurate); distil-large-v3 default hits the sweet spot for English
• Dual-engine, cross-vendor GPU — NVIDIA acceleration via CUDA (faster-whisper) and AMD / Intel / NVIDIA acceleration via Vulkan (whisper.cpp), with a CPU fallback. Eqho auto-selects the best engine for your machine — or choose it yourself in Settings → General → Inference Engine

Languages

English · Spanish · Mandarin · Japanese · Korean · Vietnamese · Arabic · Ukrainian · French · German · Portuguese · Russian · Italian

Distil models are English-optimized; for other languages pick large-v3-turbo or medium in the Models tab.

Requirements

• 64-bit Windows 10/11 (primary), Linux X11, or macOS — there is no 32-bit build; the on-device AI models require a 64-bit system
• Python 3.10+ (only when running from source)
• Optional GPU acceleration — auto-detected, no configuration:
  • NVIDIA — CUDA Toolkit 12.x for ~5× faster transcription (winget install Nvidia.CUDA)
  • AMD & Intel — no toolkit to install; the Windows build ships the Vulkan engine (whisper.cpp), which also runs on NVIDIA

Platform notes

• Linux: needs libportaudio2, xclip, and an AppIndicator library for the tray; global hotkeys and text injection target X11 (Wayland is on the roadmap).
• macOS: grant Eqho Accessibility and Input Monitoring in System Settings → Privacy & Security; first launch of the unsigned app is right-click → Open.
• Unsigned Windows binaries show a SmartScreen prompt on first run — "More info" → "Run anyway". Code signing is planned.

FAQ

Is my audio uploaded anywhere? No. Transcription runs locally via CTranslate2. The only network access is the one-time model download.

Does it cost anything? No. Free, open source, no premium tier.

Does it work on AMD or Intel GPUs? Yes. Besides the NVIDIA (CUDA) engine, Eqho ships a cross-vendor Vulkan engine (whisper.cpp) that accelerates dictation on AMD Radeon and Intel Arc / Iris graphics — and NVIDIA too. Eqho detects your hardware and picks the fastest engine automatically; you can also set it manually under Settings → General → Inference Engine.

How accurate is it? Whisper-class accuracy — state of the art for open speech recognition. Pick a bigger model for tougher audio.

Where are my settings and history? In your user config folder (%APPDATA%\Eqho on Windows) — never in the cloud.

Configuration & docs

Settings persist to settings.json in your config folder — everything is also adjustable from the tray menu and dashboard. For development docs, see AGENTS.md, docs/ai/, ROADMAP.md, and CHANGELOG.md. The repo is CodeGraph-indexed (codegraph init after cloning).

Under the hood

For the technically curious (the website stays non-technical on purpose):

• Two interchangeable inference engines, selected automatically or in Settings → General:
  • faster-whisper on CTranslate2 — CUDA on NVIDIA, optimized INT8/FP16 on CPU
  • whisper.cpp via pywhispercpp — Vulkan acceleration on AMD Radeon, Intel Arc/Iris and NVIDIA, from one wheel (built by tools/mkwheel.py)
• Subprocess model host — the model runs in a separate process (src/model_host.py), so engine/model switches never freeze or crash the UI, and VRAM is released cleanly
• Adaptive VAD with a self-calibrating noise floor and user sensitivity; audio is captured with sounddevice/PortAudio at 16 kHz mono
• Deterministic text pipeline (src/textproc.py) — casing/spacing cleanup, voice commands, user replacements; never rephrases what you said (no LLM touches your words)
• UI: tkinter + CustomTkinter dashboard, a frameless always-on-top overlay with a numpy-rendered animated "listening pill" (src/pillfx.py), tray via pystray
• Injection: clipboard-paste or simulated typing per app, with per-app rules
• Packaging: PyInstaller bundles + Inno Setup installer (Windows), tar.gz/AppImage (Linux), dmg (macOS) — all built by CI on every version tag (.github/workflows/release.yml)
• Python 3.10+, ~zero telemetry: the app never phones home; the only network call is the model download you ask for

License

Eqho is licensed under the GNU AGPL-3.0: free to use, study, and modify — if you redistribute it or a derivative (including as a network service), you must keep the copyright and attribution notices and release your version under the same license. © Daniel Mevit.

Made by damt.xyz — freelance design & development.

Provided "as is", without warranty of any kind.

---

_{Keywords: speech to text · voice to text · dictation software · offline speech recognition · free dictation app · voice typing for Windows · OpenAI Whisper app · whisper.cpp GUI · faster-whisper GUI · private speech to text · local voice recognition · no-cloud transcription · talk to type · voice keyboard · speech recognition without internet · dictate into any app · Dragon NaturallySpeaking alternative · Windows Voice Access alternative · open source dictation · GPU accelerated transcription · CUDA speech to text · Vulkan speech to text · AMD GPU dictation · AMD Radeon voice to text · Intel Arc dictation · Intel GPU speech recognition · cross-vendor GPU transcription · Whisper on AMD · runs on any GPU · real-time transcription overlay · hands-free typing · accessibility voice input · multilingual dictation}