Inspiration

I frequently have tense conversations with my family (basically arguments), but afterwards, its always hard to review constructively, so we keep staying in the same conversation pattern. I wanted a private, practical way to revisit the actual words that were said without turning it into a diagnosis, a verdict, or a guess about someone’s intent.

What it does

Speakwise turns an authorized conversation recording into a speaker-labelled, timestamped transcript and a concise coaching review. It identifies evidence-supported communication patterns, links every observation to the exact quote and time, and highlights one strained moment with a respectful, goal-preserving rephrase.

How we built it

Speakwise is a TypeScript and Express web app built with Codex using GPT-5.6 Terra.

The live path uses AssemblyAI for diarized transcription and GPT-5.6 through the OpenAI Responses API for a schema-constrained coaching report. Before an insight is shown, the app validates its quote, speaker, and timestamp against the normalized transcript; unsupported output is dropped.

The product includes an explicit authorization gate, a synthetic no-key demo for reliable testing, deletion controls, sanitization, and a local daily cap for live provider runs.

Testing

Open https://speakwise-vmf9.onrender.com. Choose Try the synthetic demo or "upload a recorded conversation" to exercise the full flow.

Built With

Share this project:

Updates