VoiceBridge

Bilingual Korean+English AI voice memo assistant. Record a memo in Korean, English, or both — get a transcript with per-segment language tags, an LLM summary, extracted action items you can add to Google Calendar, and a semantic search bar to ask questions across all your past memos.

Live demo: https://voicebridge-one.vercel.app

Features

• Bilingual STT — Gladia transcribes Korean, English, and code-switched audio (ko+en with code-switching enabled)
• Speaker diarization — transcript segments are grouped by speaker, with inline speaker renaming
• Per-segment language tagging — each transcript segment is tagged [KO] or [EN] with one-tap translation
• AI summary + action items — Claude Sonnet extracts a summary and concrete TODOs via structured tool use
• Google Calendar integration — one-click to add any action item as a calendar event
• Semantic search — ask questions across all your memos; answers are grounded in retrieved excerpts with citations (RAG)
• Eval harness — WER, LLM-as-judge, action item precision/recall, retrieval P@3/MRR — CI blocks regressions

Stack

Layer	Choice
Frontend	Next.js 16 (App Router) + TypeScript + Tailwind + shadcn/ui
STT	Gladia API (pre-recorded, diarization + code-switching) — migrated from OpenAI Whisper API, which has no diarization
LLM	Claude Sonnet 4.6 (Anthropic SDK, prompt caching)
Embeddings	OpenAI text-embedding-3-small
Vector store	Upstash Vector
Database	Supabase Postgres
Audio storage	Vercel Blob
Auth	Clerk
Calendar	Google Calendar API v3
Deploy	Vercel

Setup

1. Clone and install

git clone https://github.com/thecyprince/voicebridge
cd voicebridge
npm install

2. Set environment variables

cp .env.example .env.local
# Fill in all values — see .env.example for instructions

You'll need accounts for: Gladia, OpenAI, Anthropic, Upstash, Supabase, Vercel, Clerk, and Google Cloud (Calendar API).

3. Create the Supabase table

Run supabase/schema.sql in the Supabase SQL editor.

4. Run locally

npm run dev
# Open http://localhost:3000

Evals

# Populate evals/data/ with audio files + ground truth first (see evals/*.ts for format)
npm run evals

Note: evals/wer.ts still measures OpenAI whisper-1, not the production Gladia STT — the WER targets below are from the Whisper era.

Eval	Target
STT WER (English)	< 10%
STT WER (Korean)	< 15%
STT WER (code-switched)	< 20%
Summary faithfulness	≥ 4.5/5
Summary relevance	≥ 4.0/5
Action item precision	≥ 0.90
Action item recall	≥ 0.70
Retrieval P@3	≥ 0.70
Retrieval MRR	≥ 0.60

Deploy

vercel --prod
# Set all env vars in the Vercel dashboard