Open Source & Free AI Talker & Transcriber — text ⇄ speech, both ways, 100% local
Select text anywhere. Hit a shortcut. Put your headphones on. Or hit the other shortcut, speak — and your words land at the cursor, punctuated.
Quantum Talker is a free, open-source desktop companion (macOS · Windows · Linux) that reads any text out loud using local, offline neural voices — and turns speech back into text with NVIDIA Parakeet running on your own CPU. Dictate into any app, or drop audio files for clean, punctuated transcripts. No cloud, no account, no telemetry, no subscription. Your words never leave your machine — in either direction.
Built in a dark-first, warm design language with a full light mode.
Two more ways to get audio into the app, both transcribed on your own machine.
Apple Voice Memos (macOS): your recordings, listed in-app. Connect the folder once, and every memo sits one click away from a clean, punctuated transcript.
Telegram voice notes: record on your phone, send it to a bot you own, and the text comes back in the chat. Off until you switch it on. Setup guide →
Both land in your speaking history, and a new transcript viewer opens any of them full-screen: read the whole thing, or split it into sections you can copy one at a time.
- Global shortcut (default
⌘⇧Space/Ctrl+Shift+Space) — captures the selected text in whatever app you're in and starts reading it. - Streaming playback — speech starts after the first sentence is synthesized (with a buffering indicator for the first moments) while the rest generates in the background.
- Floating mini player — an always-on-top pill with a live waveform timeline (click to seek), ±15 s skips, playback speed (0.8×–2×), pause/stop, and a drag grabber. On macOS the menu-bar icon opens the same player as a dropdown panel.
- Minutes-saved dashboard — tracks how much reading time you've reclaimed by listening (benchmarked against the ~238 wpm average silent-reading speed).
- Markdown-aware speech —
#,**,*, links, code marks, and tables are never pronounced; headings get natural intonation and real "breather" pauses in the audio. - 28 voices, female and male (American + British English), quality-graded, one-click previews.
- Speaking pace in words-per-minute (100–280 wpm) with estimated listening time, plus an instant playback-speed multiplier during playback.
- Export to MP3 or WAV — from the current read (instant) or from any history entry.
- History — everything you've read is retrievable: re-read, export, copy, delete.
- Media keys, dock playing indicator, launch-at-login, dark/light/auto themes.
- Privacy page — a plain-language explanation of exactly what the app does and doesn't do.
- Dictate anywhere (default
⌘⇧D/Ctrl+Shift+D) — one hotkey starts the take, the same hotkey stops it. Your transcript is copied and pasted right at your cursor, with punctuation and capitalization, after a soft success chime. - The pill goes live while you speak — pulsing red, a real-time waveform of your voice, elapsed time, and the name of the microphone it's listening to (also shown on hover).
- Drop audio files (mp3, wav, m4a, flac, ogg, opus…) anywhere on the window — each becomes a clean transcript you can copy, or hand straight back to the reader to hear.
- Pick your input — System Default or any microphone, from Settings, from the Scribe tab, or straight from the tray menu; the recording pill always states which one is live.
- Tray extras — Start/Stop Dictation and Copy Last Transcript live in the menu-bar / tray menu.
- Speaking history vs reading history — everything you dictate or transcribe is kept separately from what the app read to you; filter with one click, copy or re-read any take.
- Typing-time-saved dashboard — dictating at ~150 wpm beats typing at ~40 wpm; Scribe counts what that saves you.
Your Apple Voice Memos, listed right in the app. Connect the recordings folder once (macOS shows a folder picker already pointed at it) and every memo appears in the Memos tab with its title, date, and length. One click transcribes it locally; the transcript is ready to copy and lands in your speaking history. Already-transcribed memos show as such across restarts. Prefer zero setup? Dragging memos out of the Voice Memos app onto the window works too.
Make your own Telegram bot (two messages to @BotFather), paste its token into Settings, and every voice note or audio file you send that bot gets fetched, transcribed on your machine, and filed in your history. By default the bot replies in the chat with the text, so notes you record on your phone come back to you as words. Off by default, token stored encrypted via the OS keychain, and notes sent while the app is closed catch up on the next launch.
This is the one feature that talks to a server. While it's on, the app contacts
api.telegram.org and nothing else: it fetches the audio you sent and posts the reply
back. Transcription still happens entirely on your machine, and the bot token is
encrypted with the OS keychain and never reaches the UI. Worth saying plainly, though:
a chat with a Telegram bot is ordinary cloud messaging, so give it the same trust you
give any Telegram chat.
Any transcript opens full-screen and selectable, either in one piece or split into sections at sentence boundaries, each with its own copy button. Reachable from History, from Memos, and from the Scribe tab.
The built-in engine is Kokoro-82M (Apache-2.0) running via ONNX in a dedicated utility process — the best quality-per-megabyte in open TTS today. The quantized model (~86 MB) downloads once on first launch and is cached locally; after that the app works fully offline. Voice embeddings ship with the app.
Dictation and file transcription run on NVIDIA Parakeet TDT 0.6B v2 (CC-BY-4.0) — the model that topped the Open ASR leaderboard — as int8 ONNX through sherpa-onnx in its own utility process. It decodes faster than 30× real-time on CPU and emits real punctuation and capitalization, so what lands at your cursor reads like you typed it. The model (~660 MB) downloads once, on first use (never at install), then everything is offline. A multilingual variant (Parakeet TDT 0.6B v3, 25 European languages) is one click away in Settings.
| Model | Direction | Languages | License | Status |
|---|---|---|---|---|
| Kokoro-82M (hexgrad) | speaks | EN-US, EN-GB | Apache-2.0 | Built-in |
| Parakeet TDT 0.6B v2 (NVIDIA) | listens | EN | CC-BY-4.0 | Built-in |
| Parakeet TDT 0.6B v3 (NVIDIA) | listens | 25 | CC-BY-4.0 | Built-in (Settings) |
| Piper (Rhasspy) | speaks | 47 | MIT | Roadmap |
| Chatterbox (Resemble AI) | speaks | 23 | MIT | Roadmap |
| XTTS-v2 (Coqui) | speaks | 17 | CPML | Roadmap |
| Qwen3-TTS (Alibaba) | speaks | 10 | open weights | Roadmap |
Grab the latest release for your platform from the Releases page:
- macOS —
.dmg(Apple Silicon) - Windows — NSIS installer (x64)
- Linux —
.AppImage(x86-64)
macOS note: reading the selection — and pasting a finished dictation at your cursor — requires the Accessibility permission (one synthetic ⌘C / ⌘V, nothing else). Dictation additionally asks for the Microphone permission the first time you use it. The app guides you through both, once.
npm install
npm start # run in dev
npm test # unit tests (node:test)
npm run coverage # coverage report (c8) — 100% on the core logic modules
npm run dist # build installers for mac + win + linuxRequires Node ≥ 20 for tooling; the app runs on Electron's bundled runtime.
src/main/index.js app shell: windows, tray, shortcuts, dictation state machine, IPC, stores
src/main/engine.js TTS host in a utility process: kokoro-js (ONNX), sentence-streamed
PCM, markdown-aware pacing, WAV/MP3 encoding
src/main/asr.js STT host in a utility process: Parakeet TDT via sherpa-onnx,
resumable model download, silence-aware chunked decoding
src/main/asr-models.js Parakeet preset catalog + download planning (pure, tested)
src/main/audio.js PCM helpers: resample, downmix, silence-snapped chunking (pure, tested)
src/main/transcript.js transcript tidying + audio-file gatekeeping (pure, tested)
src/main/history-util.js read vs spoken history bookkeeping (pure, tested)
src/main/telegram.js Telegram poller in a utility process: long-poll getUpdates,
file download, transcript replies, offline backoff
src/main/telegram-util.js update classification, dedup ring, chat pairing, message
splitting (pure, tested)
src/main/telegram-host.js bot-token lifecycle: safeStorage encryption, poller supervision
src/main/voicememos.js Voice Memos folder scan + CloudRecordings.db title lookup,
tolerant of a missing or renamed database
src/main/voicememos-util.js memo listing, id derivation, transcribed-state merge
(pure, tested)
src/main/keysim.js per-platform copy/paste keystroke commands (pure, tested)
src/main/markdown.js markdown → speech segments with breather pauses
src/main/selection.js selected-text capture + paste-at-cursor (native ⌘C/⌘V helper)
src/preload/index.js contextBridge API
src/renderer/ zero-framework ES-module UI; the floating pill doubles as
mini player and live dictation recorder (waveform + mic capture)
Quantum Talker is free software, released under the MIT license — use it, read it, fork it, improve it. See LICENSE.
All synthesis — and all listening — happens on your machine. © 2026 Adam Blvck