No description
  • Rust 98.8%
  • Python 0.7%
  • PowerShell 0.3%
  • Shell 0.2%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-11-18 17:00:43 +10:30
assets eaSpeak: push-to-talk dictation for Windows with local GPU speech recognition 2026-11-03 20:26:18 +10:30
crates Show live meeting captions and preserve vocabulary spellings 2026-11-18 17:00:43 +10:30
docs Show live meeting captions and preserve vocabulary spellings 2026-11-18 17:00:43 +10:30
scripts Start the tray app through the desktop shell, so autostart really works 2026-11-04 08:53:03 +10:30
.gitignore eaSpeak: push-to-talk dictation for Windows with local GPU speech recognition 2026-11-03 20:26:18 +10:30
AGENTS.md Remove duplicated logic and dead code found by the technical review 2026-11-10 18:31:46 +10:30
Cargo.lock Meeting mode: transcribe a whole meeting as it happens, with notes and attachments 2026-11-10 11:55:34 +10:30
Cargo.toml eaSpeak: push-to-talk dictation for Windows with local GPU speech recognition 2026-11-03 20:26:18 +10:30
CLAUDE.md eaSpeak: push-to-talk dictation for Windows with local GPU speech recognition 2026-11-03 20:26:18 +10:30
LICENSE eaSpeak: push-to-talk dictation for Windows with local GPU speech recognition 2026-11-03 20:26:18 +10:30
README.md Show live meeting captions and preserve vocabulary spellings 2026-11-18 17:00:43 +10:30
rustfmt.toml eaSpeak: push-to-talk dictation for Windows with local GPU speech recognition 2026-11-03 20:26:18 +10:30
THIRD_PARTY_NOTICES.md eaSpeak: push-to-talk dictation for Windows with local GPU speech recognition 2026-11-03 20:26:18 +10:30

eaSpeak

Hold Right Alt, speak, let go, and the text appears where you were typing. It also stays on the clipboard. English, Chinese and 50 more languages, mixed in one sentence too ("这个 bug 我们 fix 了吗"). The speech is recognised on your own GPU, nothing leaves the PC, and recordings are never saved.

eaSpeak is the Windows counterpart of eaSnippet's dictation on the Mac, with the same philosophy:

  • Local and private. Qwen3-ASR runs on your GPU through llama.cpp. Audio lives in memory only.
  • Native and lean. It's a 3 MB Rust program that uses Win32 directly: a low-level keyboard hook, WASAPI, UI Automation, and Direct2D for the UI. Idle, it uses no CPU at all (measured below): no timers, no polling, and the speech server is suspended between dictations.
  • Command line at the core. Every setting, model and dictation can be driven with easpeak, and --json output makes it easy for an AI agent to manage it. There is an MCP server too.
Dictate Hold Right Alt (configurable) for ¼ s, speak, release. A pill at the bottom of the screen shows a mic that pulses with your voice and your words as you speak, then "Inserted". Tapping Right Alt, or using it in a shortcut (Alt+C), never starts a dictation. Pressing any other key while talking cancels.
Where text goes Always onto the clipboard. Typed (pasted) into the focused text field when there is one. Never into a password field; buttons, lists or the desktop get the clipboard only.
Model manager Tray icon › Speech Models…, or easpeak models. Download, switch or remove models, and install the GPU engine.
Clean-up CJK/Latin spacing ("比较 focus 在"). In English and Chinese it also drops "um"/"uh" and stumbles ("the the"), and writes numbers as digits ("port 8080"). Ported test-for-test from eaSnippet.

Install

eaSpeak is built from source in WSL (or any Linux) and installs itself on the Windows side. You don't need Visual Studio.

cargo install cargo-xwin --locked                 # once; also needs LLVM (clang, lld, llvm-rc)
scripts/deploy.sh                                 # build, install or update, start
easpeak doctor                                    # what's left to set up
easpeak runtime install                           # llama.cpp for your GPU (CUDA 13 on NVIDIA: 593 MB)
easpeak models pull qwen3-asr-1.7b                # the speech model (2.5 GB)

deploy.sh copies easpeak.exe and easpeak-app.exe to %USERPROFILE%\.easpeak\bin, adds that folder to PATH, creates a Start menu entry, turns on start with Windows (pass --no-autostart to skip it) and starts the tray app. It also links ~/.local/bin/easpeak in WSL. Running it again updates eaSpeak: the app is stopped, the programs are swapped and the app is started again. The model manager window opens by itself the first time, if anything still needs downloading.

Everything lives in %USERPROFILE%\.easpeak (config.toml, models\, runtime\, logs\, bin\). To remove eaSpeak, run easpeak uninstall --yes (add --purge to delete the models and settings too).

Command line

easpeak status                       # running? hotkey, model, GPU engine, speech server
easpeak doctor                       # every check, each with the command that fixes it
easpeak models                       # list · pull <id> · use <id> · rm <id> --yes
easpeak runtime install --backend cuda|vulkan|cpu
easpeak config list                  # every setting, its value, default and meaning
easpeak config set hotkey.key RCtrl  # the running app picks changes up at once
easpeak config set voice.vocabulary "Qwen, eaSpeak, PeoplePlanner"
easpeak transcribe meeting.wav       # uses the app's warm model; from WSL: "$(wslpath -w file.wav)"
easpeak bench bench/ascend           # speed and accuracy (a .txt next to each WAV holds what was said)
easpeak dictate start --file a.wav   # drive a dictation (agents and tests); `dictate stop` returns the text
easpeak app start|stop|restart       # the tray app
easpeak autostart on|off
claude mcp add --scope user easpeak -- easpeak mcp

For names you add to the vocabulary, dictation and meeting captions use your spelling when they hear the component words or initials. For example, “People Planner” becomes “PeoplePlanner” and “W S L” becomes “WSL”.

Every command accepts --json, and errors arrive as {"error": {"code", "message"}} on stderr. easpeak docs prints the guide for agents (docs/llms.txt).

Meeting mode

easpeak meeting start --title "Pricing with Acme"   # or tray › Start Meeting
easpeak meeting note "Ask about the deadline"
easpeak meeting attach diagram.png --caption "the old flow"
easpeak meeting screenshot
easpeak meeting stop                                 # prints where it's saved
easpeak meeting summary                              # decisions, action items, open questions
easpeak meeting export latest --format md            # or txt (raw transcription), json

Or use the meeting window (tray › Meeting Window…, or easpeak meeting window): name the meeting and start it, watch what's said appear, type notes into the box (Enter adds one; dictating into it works too), attach files with the button, by dropping them on the window or by pasting, and add screenshots with the button or by pasting a snip. It lists earlier meetings; a click opens one. Closing the window doesn't end the meeting: the tray shows it's running, and ends it.

A meeting listens to the microphone and to what the computer plays, so both sides of a call are heard. The meeting window shows a provisional Live line while someone speaks. The recogniser can offer the first update after 0.6 seconds of audio. During continuous speech it can refresh every half a second, with more time between updates when the GPU takes longer. It transcribes the full utterance at a pause for the saved timeline. Closing or minimising the window stops the extra caption work. Push-to-talk also takes priority over provisional captions. Notes, files and screenshots land in the timeline at the moment they were added. Ending the meeting writes meeting.md into its folder (.easpeak\meetings\<date-time>): the notes and attachments first, then the transcript with them in place. That document is what an agent reads to write a summary.

meeting summary has a small text model on the same GPU (Qwen3 4B, a 2.5 GB download on first use) write the summary, in the meeting's main language or the one you name: about 10 s for a short meeting, a minute for an hour. The speech model steps off the GPU while it runs and comes back after. The summary is saved beside the meeting and put at the top of meeting.md.

Measured on the RTX 3050: an utterance of 10 s is transcribed in about a second; on the mixed Mandarin-English benchmark clips played as one meeting, accuracy matches dictation's (13.4% against 12.8% mixed error rate on the same audio, one token apart). While nobody speaks, listening costs 0.2% of one CPU core. Only text is saved unless meeting.keep_audio is on. Tell the people you're meeting that it's being transcribed.

Accuracy and speed

Measured on this laptop (RTX 3050 6 GB Laptop GPU, i5-12450HX), llama.cpp b12065, using 24 utterances from ASCEND. That is spontaneous Mandarin-English conversation, 8 each of English, Mandarin and mixed, about 3.5 s each, which is much harder than deliberate dictation. The scores are mixed error rates (a Chinese character or an English word counts as one token; fillers are ignored on both sides). Times are median per utterance, with the model loaded.

Model Engine Time All English Mixed Mandarin
Qwen3-ASR 1.7B (default) CUDA 13 279 ms 13.0% 16.7% 16.0% 6.3%
Qwen3-ASR 1.7B Vulkan 283 ms 13.0% 16.7% 16.0% 6.3%
Qwen3-ASR 0.6B CUDA 13 126 ms 14.3% 17.4% 18.8% 6.6%

Long recordings run at about 5.5% of real time: 31 s of speech takes 1.7 s and 93 s takes 5.3 s. While you speak, the live preview already finishes the earlier 28 s pieces, so after you release only the last piece is left to do. Loading the model onto the GPU takes about 4 s. That happens on the first dictation after start-up, and again after 30 idle minutes, when eaSpeak frees the GPU memory; it starts as soon as the hold counts, so it overlaps with you speaking. To never wait for it, keep the model loaded: easpeak config set server.preload true and easpeak config set server.idle_minutes 0 (it then holds about 3 GB of GPU memory, but no CPU). Many of the remaining "errors" are transcription conventions rather than mistakes ("MSc" vs "m s c", "favorite" vs "favourite", and stumbles that dictation deliberately removes). See docs/research.md for why Qwen3-ASR was chosen, and what else was measured.

CPU, phase by phase (measured 3 Nov 2026 with a 31 s clip, share of one CPU core; the machine was busy with other work at the time, so recognition times are on the high side):

Phase eaSpeak app Speech server (llama.cpp)
Idle 0 0 (suspended, the model stays loaded on the GPU)
Recording, live preview off 3–5% (the pill: meter and clock) 0
Recording, live preview on 4–8% ≤ 40% (one update at a time, resting 1.5× its duration after each)
Recognising < 1% one core while the GPU works (CUDA waits for the GPU by spinning)

The model runs entirely on the NVIDIA GPU (--device CUDA0, all layers; easpeak status names the GPU). The Vulkan build waits for the GPU without spinning: about 30% less CPU while recognising, about 10% slower (easpeak config set runtime.backend vulkan). With the live preview on, letting go after a pause inserts the last update's text at once (~0.2 s) instead of recognising again.

Build and test

cargo test --workspace                       # text clean-up, protocol, audio, config and meeting captions
scripts/build.sh                             # target/x86_64-pc-windows-msvc/release/*.exe
cargo xwin clippy --target x86_64-pc-windows-msvc --workspace --all-targets
python3 scripts/fetch-bench.py               # the benchmark clips (not committed; CC BY-SA)

The layout is described in docs/ARCHITECTURE.md; AGENTS.md has the conventions for people and agents changing it.

Licence

MIT. The text clean-up is a port of eaSnippet's (MIT). eaSpeak downloads llama.cpp (MIT) and Qwen3-ASR (Apache-2.0) and doesn't redistribute them. See THIRD_PARTY_NOTICES.md.