SpeakLocal — an offline AI language tutor that never hears you online

The fastest way to learn a language is to speak it badly to somebody patient. This is that, running entirely on your phone, so the recording of you fumbling never leaves it.

Speaking is the skill nobody practises

You can finish a flashcard app's entire tree and still freeze when someone speaks to you. Reading and listening are practised constantly because they are comfortable; speaking is avoided because it is not, and because it requires a partner willing to sit through your first hundred bad sentences.

An AI tutor solves the partner problem. But the mainstream ones stream your microphone to a server, which means the sound of you struggling — and everything you happen to say — becomes somebody else's data. SpeakLocal runs the model on the phone instead.

What a session with the offline AI language tutor looks like

You pick a scenario, speak, and get a response plus feedback. Everything below serves that loop.

• Scenario conversations. Practice happens in situations rather than in the abstract: ordering in a coffee shop, asking directions, introducing yourself, and — in the Pro set — a job interview, a doctor's visit, hotel check-in, a first date, a business meeting, filing a complaint, a phone call with no visual cues, an emergency call. Each has its own register and vocabulary.
• Three languages, three tutor personas. English, Mandarin Chinese and Japanese, each with its own prompt template and locale so the tutor behaves appropriately for the language rather than being a translated English bot.
• Pronunciation scoring at phoneme level. The pronunciation trainer breaks your attempt into phonemes and scores them individually, so you find out that it is one specific sound letting you down rather than getting a vague overall mark.
• Mandarin tone detection. For Mandarin, the app extracts a pitch contour and classifies the tone shape — flat-high, rising, dip-rise, falling or neutral — then shows it as a curve against the target. Tones are the single hardest thing for most Mandarin learners and almost never get real feedback.
• A vocabulary vault with spaced repetition. Words you meet go into a vault scheduled by an SRS algorithm, reviewed with the familiar Again / Hard / Good / Easy grading, with a due-today count on the home screen.
• Progress you can see. A confidence chart, streaks and per-session summaries, plus a daily goal set during onboarding.
• A data inspector. Settings has a panel showing how much data is stored on-device, how much of it is encrypted, and how much has been sent to the cloud. Publishing that last figure is unusual, and it is the point.

A firewall, not a promise

Most privacy claims are policy. SpeakLocal implements one in code: an allowlist network guard sits in front of the HTTP client and rejects any host that is not explicitly permitted. The permitted set covers the model download hosts and the ad and purchase SDKs — nothing else can make a request, and crash reporting is only allowlisted if you opt in.

Progress and vocabulary are held in an encrypted local store, and there is an Erase all my data action that does exactly that. Your microphone audio is processed on the device and is not uploaded.

Who it is for

Intermediate learners stuck at the point where they understand far more than they can produce. Twenty minutes of daily speaking with immediate correction moves that faster than another month of vocabulary drills.

Mandarin learners specifically, because of the tone feedback. If you have been told your tones are wrong but nobody can explain how, seeing your pitch contour drawn against the target is the missing piece.

And anyone preparing for a specific event — a job interview in a second language, a medical appointment abroad, a business meeting — who wants to rehearse that exact conversation a dozen times without an audience.

Privacy and requirements

The language model is downloaded once during onboarding from a public model host, after which conversation runs on the device. Speech recognition and text-to-speech use the platform's own engines. Audio is processed locally and not uploaded. Data lives in encrypted local storage and can be erased from Settings.

The app is free with a Pro purchase that unlocks the wider scenario set. It is a practice tool, not a certified assessment.

Frequently asked questions

Is my voice sent to a server? No. Speech is processed on the device and the tutor model runs locally. An allowlist network guard rejects requests to any host not explicitly permitted.

Which languages can I practise? English, Mandarin Chinese and Japanese, each with its own tutor persona and locale.

How does the pronunciation feedback work? Your attempt is scored per phoneme, so you can see which specific sound is letting you down. For Mandarin, the app also classifies your tone from the pitch contour and draws it against the target shape.

Does it work without an internet connection? Yes, once the model has been downloaded during setup. Conversation, scoring and vocabulary review all run offline.

Can I delete everything? Yes. Settings has an Erase all my data action, plus a data inspector showing what is stored on-device, how much is encrypted, and how much has been sent to the cloud.