TsubuTalk: AI Speaking Coach

■ TsubuTalk — AI Speaking Coach, on your device.

TsubuTalk turns the world's most capable AI chatbots (ChatGPT, Gemini, Claude) into your personal speaking coach. Record a short utterance ("tsubuyaki"), get instant 4-axis scoring on device, and then hand the transcript off to your favorite AI with a built-in coaching prompt for in-depth feedback on logic, vocabulary, persuasiveness, or grammar.

Your recordings and transcripts stay on your phone — nothing is sent to external servers, ever.

■ Why ChatGPT / Gemini / Claude as your coach?

Whether it's English conversation, TOEFL / IELTS speaking, a presentation, or an interview, you can ask a top-tier AI to "review my logic," "improve my vocabulary," or "make this more persuasive" and get a structured critique in seconds.

TsubuTalk never calls AI APIs directly. You pick a provider (ChatGPT / Gemini / Claude) and the app hands off your transcript via Share Intent, clipboard, or browser — you paste it into the official app or site yourself. No API key needed.

- 5 built-in coaching prompts: logic / vocabulary / persuasiveness / grammar / custom
- Switch the target provider in settings (ChatGPT default, Gemini beta, Claude beta)

■ Instant on-device 4-axis scoring

Before you hand off to the AI coach, see objective scores instantly on the device:

- Volume (dBFS): average input level vs. a comfortable target range, as a 5-step badge.
- Speech rate: mora/min (Japanese) or syllables/min (English), from the transcript.
- Filler rate: detects fillers ("um", "uh" / "えーと", "あの") and their share of total tokens.
- Pronunciation clarity: a 0-to-1 score from whisper.cpp's recognition confidence.

When any axis falls outside its target range, a short text hint is shown.

■ 330 built-in topic prompts

Beyond a "freestyle" mode, TsubuTalk ships with 330 prompts across 5 locales.

- DAILY: 150 (30 x 5 locales)
- SPEECH (structured / persuasive): 150 (30 x 5 locales)
- ENGLISH (speaking drill): 30 (shared across locales)

A random prompt is offered each time so you never run out of things to say.

■ Progress dashboard

Track scores over 7 / 30 / 90 days with line charts and a 4-axis radar. Streak counter, weekly/monthly counts, and rule-based insights at a glance.

■ Mic calibration

A 5-second calibration personalizes the target dBFS range to your device, so the volume score fits your setup.

■ Privacy — recordings never leave your device

- Recordings, transcripts, scores, and calibration data stay on the device.
- Device backup is disabled (`allowBackup=false`). Data migration is JSON export via SAF only.
- Crashes are not auto-reported; manual export only.
- The only data leaving the device: AdMob ads, the first-time model download, and text you explicitly send to an external AI chat.

See the in-app privacy policy for full details.

■ Pricing — completely free, ad-supported

The app is completely free. A banner appears on the feedback screen and an interstitial during analysis (Google AdMob, personalized ads disabled). No ad-free plan, no in-app purchase, no subscription.

■ Requirements

- Android 11 or later (minSdk = 30)
- ~600 MB free storage for the speech-recognition model
- Wi-Fi recommended for the first launch (model download)

■ Supported languages (5 locales)

UI: Japanese / English / Simplified Chinese / Traditional Chinese / Korean
Speech recognition: Japanese / English / Simplified Chinese / Traditional Chinese / Korean / Auto-detect (multilingual whisper model)

■ Who is this for?

- Anyone who wants AI (ChatGPT / Gemini / Claude) to review their speech without transcribing by hand
- Privacy-conscious users who don't want recordings sent to external servers
- Learners tracking objective metrics (speech rate, filler rate, volume)