Free · Open source · On-device
Speak English like you mean it.
Engclaw scores how you pronounce a sentence, phone by phone, and shows exactly which sounds to fix. No course fee. No account. It never sighs at your “th”.
- macOS
- Windows
- Linux
- English & Tiếng Việt
What is Engclaw?
A free, open-source English pronunciation trainer that runs on your computer.
Engclaw is a free, open-source desktop app for macOS, Windows and Linux that scores how you pronounce an English sentence, phone by phone, and shows which sounds, words and intonation to fix. It runs speech models locally, so there is no account, no server and no telemetry.
Last updated
- Type
- Desktop app (Electron), open source
- Price & license
- Free · GPL-3.0-or-later
- Platforms
- macOS (Apple Silicon), Windows x64, Linux x64
- Scoring
- 30% acoustic similarity, 40% correct sounds, 30% correct words
- Privacy
- On-device; no account, no telemetry
- Interface
- English and Vietnamese
Everything in one app
Five ways to fix your accent.
-
Practice
Read a sentence aloud or speak freely. See which sounds, words and intonation to fix, with pitch curves against a reference voice.
-
Shadowing
Paste a YouTube link, repeat each line, and get scored. Switch to dictation when you want to test your ears.
-
Pronunciation game
Clear chapters one word at a time and collect stars. A try only counts at 50+ with no flagged word.
-
Dictionary
IPA, roots and word families, plus an animated mouth and tongue for each sound. Study with flashcards.
-
Converse
Optional AI speaking partner through a provider you configure. Say a sentence, get a spoken reply.
Under the hood
It listens to sounds, not guesses.
Engclaw never asks a cloud service “does this sound OK?”. A stack of models on your computer compares what you said with what you should have said.
- 1
Which sounds did you make?
A wav2vec2 phone recogniser writes down the IPA phones it heard, with no idea what you meant.
- 2
Which should you have made?
espeak-ng turns the target sentence into its expected phones.
- 3
Compare
The sequences are aligned and every mismatch is marked, giving a phoneme error rate and a list of words to fix.
- 4
Which words did you say?
A second model transcribes your speech for the word error rate. A different word is a different word.
- 5
Do you sound like it?
Kokoro says the sentence; dynamic time warping compares frames, so slower or faster is fine but different sounds are not.
- 6
Intonation
Pitch and energy curves are drawn for you and the reference.
Your score, out of 100
Fair warning: it’s a tough grader. A swallowed vowel is a wrong vowel, and one mumbled word still gets a red mark.
Models Engclaw uses
Downloaded once from Hugging Face on first use (with your consent), checksum-verified, and run locally at fp32 precision.
| Model | Used for | Size | License |
|---|---|---|---|
| wav2vec2-lv-60-espeak-cv-ft | Phone recognition: pronunciation and game | ~1.26 GB | Apache-2.0 |
| wav2vec2-large-960h | Speech recognition: free speaking and Converse | ~1.26 GB | Apache-2.0 |
| Kokoro-82M | Reference voice: “Hear it” and examples | ~0.33 GB | Apache-2.0 |
Privacy by design
Your voice stays on your machine.
No server, no account, no telemetry. Models download once from Hugging Face, are checksum-verified, and run locally on CPU or WebGPU.
No accounts
Open the app and start talking.
Local storage
Recordings, history and settings live in your app data folder.
Network on request
Only when you use YouTube, illustrations, updates or your own AI provider.
Open source
GPL-3.0 code you can read, audit and improve.
Honest requirements
- macOS
- Apple Silicon only
- Windows
- x64
- Linux
- x64 · AppImage or .deb
- Disk
- ~2.6 GB for models
- Memory
- 8 GB recommended
- Language
- English scoring · EN & VI interface
How to start practising in 3 steps
Download and install
Pick the installer for your system: macOS Apple Silicon .dmg, Windows x64 .exe, or Linux AppImage / .deb.
Approve the model download
On first use of a scoring feature, Engclaw lists each model’s name, size and license and asks before downloading about 2.6 GB.
Read a sentence aloud
Open Practice, record yourself, and see your score out of 100 with the sounds and words to fix.
Questions, answered.
What is Engclaw?
Engclaw is a free, open-source desktop app for macOS, Windows and Linux that scores English pronunciation phone by phone and shows which sounds, words and intonation to fix. It runs speech models locally with no account, server or telemetry.
Which AI models does Engclaw use?
Three local models: wav2vec2-lv-60-espeak-cv-ft for phone recognition, wav2vec2-large-960h for speech recognition, and Kokoro-82M for the reference voice. All are Apache-2.0 licensed and downloaded once from Hugging Face.
Can I practise with YouTube videos?
Yes. Shadowing mode lets you paste a YouTube link, repeat each line of the transcript and get scored, or switch to dictation. Fetching YouTube captions and audio may be against YouTube's Terms of Service, so it is intended for personal, educational use.
Does Engclaw support languages other than English?
Pronunciation scoring is English only. The app interface is available in English and Vietnamese.
Is Engclaw really free?
Yes. It is free and open source under GPL-3.0-or-later. No account, subscription or telemetry.
Does my voice leave my computer?
No. Scoring runs locally. Recordings, history and settings stay in your app data folder. Network requests only happen for features you choose to use.
How is the score calculated?
30% acoustic similarity to a reference voice, 40% correct sounds and 30% correct words.
Which platforms are supported?
macOS on Apple Silicon, Windows x64, and Linux x64 (AppImage or .deb). Intel Macs are not supported.
How much space and memory do I need?
About 2.6 GB of disk for the speech models plus the app. 8 GB of RAM is a practical starting point.
Does it work offline?
Yes, after the models are installed. Practice and the game work offline. YouTube, illustrations, updates and the AI agent need a connection.
Couldn’t afford a course? Neither could we.
Download Engclaw and start fixing your accent today.