DentBridge AI

An AI dental interpreter that converts speech into bilingual text and audio, cross-checking critical terms so Japanese dentists and international patients can communicate more safely.

Comment

Inspiration

Dentistry is not an abstract use case for me. Some of my customers are dentists, so before building this project I asked them what communication problems actually occur when an international patient visits a Japanese dental clinic.

Their answers were more specific than “we need translation.” They had seen existing translation apps struggle with dental and medical terminology, including tooth locations, symptoms, allergies, medications, anesthesia, and dosage. A small mistake—changing “left” to “right,” dropping a negation, altering a number, or confusing a tooth with another body part—can become a real safety risk.

They also explained that a normal mobile-app interaction does not match the dental environment. The dentist may be wearing gloves and working with both hands. The patient may be reclined with their mouth open and unable to touch a screen or speak normally. Dental drills and suction devices create loud, continuous noise that can trigger or confuse ordinary speech translation.

Those conversations changed the design of the project. I did not want to build another general-purpose translation screen. I built DentBridge AI to explore a safer, hands-free communication bridge designed around the actual workflow of Japanese dental professionals and international patients. It is strictly a translation tool. It does not diagnose conditions, recommend treatment, or make medical decisions.

What it does

DentBridge AI is a responsive Web/PWA application designed for iPad, Android tablets, and desktop browsers.

Before a session, clinic staff select the patient’s language and confirm the bilingual privacy and AI-translation notice. Japanese is fixed as the clinic language, while the current MVP supports Chinese, English, Korean, and Vietnamese for patients.

After one tap on Start Session, the app:

  1. Calibrates itself to the room’s background noise.
  2. Listens automatically using local voice activity detection.
  3. Detects when a person starts and finishes speaking.
  4. Uploads only the completed sentence rather than continuous room audio.
  5. Transcribes the original speech.
  6. Determines whether the speaker used Japanese or the selected patient language.
  7. Translates in the correct direction.
  8. Checks critical medical information.
  9. Displays the original text and verified translation.
  10. Speaks only a translation that passed the safety checks.
  11. Automatically resumes listening after playback.

DentBridge AI also separates normal conversation from active treatment.

Conversation Mode supports hands-free, bidirectional, free-form communication before and after treatment.

Treatment Mode is designed for periods when the patient cannot speak normally and dental equipment creates heavy noise. Before entering this mode, the app explains a stop signal in both languages. The patient practices raising a movable hand, and the dentist or assistant must visually confirm it. Only after confirmation does Treatment Mode begin.

During treatment, arbitrary background speech is not sent through generative translation. The app recognizes only a reviewed library of fixed dental phrases such as “Please open your mouth,” “Are you in pain?” and “Would you like a short break?”

The app does not use a camera to detect the patient’s hand. The dentist or assistant remains responsible for watching the patient and stopping treatment when the agreed signal is given.

How I built it

I built DentBridge AI using TypeScript, React, Next.js, and PWA technologies, with Codex as my primary engineering collaborator.

The browser uses the Web Audio API and an AudioWorklet for local voice activity detection. It measures ambient noise, detects likely human speech, ignores very short sounds, and separates each utterance after a period of silence. Microphone capture enables echo cancellation, noise suppression, and automatic gain control.

Each completed sentence is processed with gpt-4o-transcribe. Japanese and the selected patient language can be evaluated in parallel so the app can select the more reliable transcription candidate.

The confirmed text is then sent to gpt-5.6 using Structured Outputs. The model must return the detected language, source text, translation, confidence, and critical entities such as:

  • Body part and symptom
  • Tooth location and tooth number
  • Left and right direction
  • Negation
  • Medication and allergy information
  • Dosage, numbers, and dates

A second safety layer compares the source and translation. I combine model-based review with deterministic checks that protect anatomy, direction, negation, numbers, medication, allergies, and dosage.

Only an approved translation receives a short-lived, server-signed playback authorization. This prevents an unchecked translation from accidentally reaching text-to-speech.

By default, the app uses native device speech synthesis to reduce latency. During playback, microphone processing is paused so the app does not recognize and translate its own voice.

The complete hands-free state machine is:

calibrating → listening → recording → transcribing → translating → verifying → speaking → cooldown → listening

Challenges I ran into

The most difficult problem was determining when a person had actually finished speaking. A short silence may be the end of a sentence or simply a natural pause. Ending too early cuts off words, while waiting too long makes the conversation feel slow.

The clinic conversations had warned me about dental noise, and device testing confirmed that it was not a minor edge case. Dental drills, suction equipment, impacts, room conversations, and iPad automatic gain control can all resemble speech. I added noise calibration, minimum speech duration, signal-variation checks, pre-roll audio, and conservative filtering of low-confidence or repetitive transcription results.

Mobile Safari created additional challenges. Microphone capture and speech playback can compete for the iPad audio session, change perceived volume, or fail to report that playback has finished. I added playback recovery, echo-loop prevention, and safe automatic return to listening.

Medical verification also produced unexpected linguistic edge cases. Polite expressions may contain characters that look like negation, and Japanese expressions containing numbers do not always refer to tooth numbers. The checks therefore had to become language-aware without weakening protection of genuine medical facts.

Finally, safety and speed naturally compete. An independent second model review improves confidence but increases latency. I created clearly labeled safe and fast profiles while keeping deterministic medical checks in both.

Accomplishments that I am proud of

I am proud that the main workflow is genuinely hands-free. After starting the session, the dentist and patient can complete multiple exchanges without pressing or holding buttons.

Other accomplishments include:

  • Automatic Japanese/patient-language direction detection
  • Local VAD that avoids uploading continuous silence
  • Prevention of rejected translations from reaching TTS
  • Protection of anatomy, negation, direction, tooth numbers, dosage, medication, and allergies
  • Echo-loop prevention during speech playback
  • A treatment mode using only reviewed fixed phrases
  • A bilingual, practiced hand-signal confirmation before treatment
  • No camera-based gesture surveillance
  • Browser-memory-only conversation storage
  • Automatic recovery from common network and iPad audio failures
  • Mock scenarios for translation, rejection, treatment mode, and reconnection
  • 119 passing automated tests, plus passing lint, type checking, and production builds

DentBridge AI is still a field-test MVP, not a clinically certified medical device, but it demonstrates how hands-free translation can be combined with explicit safety gates.

What I learned

I learned that a medical translation tool should be judged not only by what it translates, but also by what it refuses to say.

Generative models are useful for multilingual understanding, but deterministic checks remain valuable for facts that must not change. The strongest approach combines transcription, structured translation, independent review, and explicit comparison of critical entities.

I also learned that mobile audio behavior is part of the core product architecture. Echo cancellation, playback timing, microphone recovery, background transitions, and device voice selection directly affect whether a hands-free experience is usable.

Most importantly, an app cannot replace observation and communication between a dentist and patient. Technology can explain and display a stop signal, but a human must still watch for it and respond immediately.

What's next for DentBridge AI

My next step is controlled testing in realistic dental environments with masks, background conversations, suction equipment, dental drills, and external microphones.

I plan to:

  • Measure transcription accuracy, translation errors, false rejections, and end-to-playback latency
  • Expand the dental test set with professional bilingual review
  • Have qualified medical interpreters review all fixed treatment phrases
  • Tune voice activity detection for different clinic rooms and devices
  • Improve iPad speech playback and external microphone support
  • Add more reviewed treatment phrases without allowing arbitrary treatment-room translation
  • Define clear escalation procedures when the app cannot translate safely
  • Conduct privacy, medical-information, and operational compliance reviews for use in Japan
  • Perform formal usability and safety studies before production clinical deployment

DentBridge AI is not intended to replace professional interpreters or clinical judgment. Its goal is to make everyday dental communication more accessible while ensuring uncertain translations are never spoken as if they were certain.

Built With

Share this project:

Updates