English Voice Coach

I created a private Android speaking coach that accepts voice, typed sentences, and photographed text, then provides corrections, explanations, natural alternatives, playback, and repeat practice.

I designed and implemented the Flutter app, Android/JNI integration, local inference, storage, model setup, test coverage, release preparation, and practice experience.

Watch the four demo videos ↗All selected work →

Current release identity · local runtime readiness and topic-based practice

Role: Independent product · implemented end to end
Platform: Flutter · Android/JNI
Local runtime: whisper.cpp · LiteRT-LM
Release: Play closed test · in review

01 Delivery

One practice loop across voice, text, and real-world English.

Voice practice records a short phrase and transcribes it with native whisper.cpp. Text practice sends typed English directly to the correction boundary, while photo practice uses Android image selection and ML Kit OCR before entering the same correction flow. Gemma through LiteRT-LM returns a correction, a natural alternative, an explanation, and structured issue details.

Flutter owns the product interface and state. Android platform channels own recording, local model files, native inference, OCR, and speech playback. Completed attempts are organized into topic-based review. Corrected or natural phrases can be played and repeated locally, and practice content stays on the device during normal use.

Three useful input paths

Speak, type, or photograph English text. Every input goes through the same correction, practice-history, and review flow.

On-device inference and fallbacks

Whisper transcription and Gemma correction run through Android platform channels. When correction fails after a successful transcription, the transcript stays available. Android TextToSpeech handles playback when the optional neural voice is unavailable.

Word-level repeat matching

The percentage compares the target phrase with the locally transcribed repeat at word level. It shows how closely the words match; pronunciation, accent, and fluency are outside this measurement.

Tested on a physical Android device

I tested the implemented flows on a Samsung SM-S938U1. The release record includes 150 passing Flutter tests, analyzer checks, APK and App Bundle generation, signing, and Play upload.

02 Interface evidence

Current Android release-candidate interface.

The product began under the working name English Practice. The screenshots below show the current English Voice Coach release-candidate interface.

English Voice Coach voice practice screen with topic selection and a large tap-to-record microphone control — Voice practice · native recording and local transcription entry point

English Voice Coach result showing original text, correction, natural version, playback, and explanation — Correction result · corrected and natural phrasing with listening and repeat actions

English Voice Coach repeat-practice screen showing a 76 percent word-level phrase match and the transcribed repeat — Repeat practice · word-level target-to-transcript comparison

English Voice Coach review screen grouping saved practice attempts and common issues by topic — Review · saved attempts, issue categories, search, and practice-again workflow

03 Delivery boundary

Setup, privacy, and release status.

After setup, transcription, correction, phrase matching, practice history, and supported voice playback run on the Android device. First-run setup downloads validated model files over HTTPS and needs substantial storage; a feedback report containing generated output leaves the device only after the user reviews it and taps Send.
The current Android release is in Google Play closed-test review and is unavailable for public installation.