05 / LEGAL SERVICES / MEETING INTELLIGENCE

A transcription engine that joins the call, translates as it listens, and never loses a word even when the connection drops.

Client: Premier legal firm, UK & Dubai

Tags: Legal Services · Real-Time Transcription · Multilingual · LLM Correction · Encrypted Data

// Problem

Cross-border legal meetings mix languages and highly specific terminology — client names, matter names, technical and legal terms — that generic transcription tools consistently mangle. There was no reliable way to get a clean, unified English transcript of a mixed-language conversation in real time, and meeting recordings needed to be handled as sensitive data throughout, not left sitting in a plaintext database.

// The Contribution

A purpose-built transcription system that joins the meeting, transcribes and translates live, and corrects itself against that specific meeting's own vocabulary as it goes. Before the meeting, any background the team provides is parsed into a structured glossary; during the meeting, every line of transcript is checked against that glossary by an LLM correction pass that fixes names and terms without ever touching the surrounding spacing or wording. If the underlying speech connection drops, a short rolling buffer replays recent audio automatically so the transcript never has a gap. After the call, a second pass produces a cleaner, diarized final transcript — deliberately keeping mixed-language segments exactly as spoken rather than flattening them into one language. Every recording and transcript sits behind an encrypted database and API-key authentication.

// Technical Details

• Real-time audio pipeline: meeting bot → speech engine (language ID + live translation to English) → LLM correction → formatted display output
• Per-meeting glossary generation: user-supplied context parsed by an LLM into structured terms, fed into both the live correction engine and the post-meeting cleanup pass
• Correction engine preserves exact token spacing and streaming partial text — no autocomplete, no rephrasing
• Automatic reconnect on speech-engine session limits, with a rolling buffer replaying recent audio so no words are dropped
• Post-meeting pass: full audio re-transcription, multi-segment diarization and merge, optional filler/profanity cleanup — mixed-language lines explicitly preserved, never auto-translated
• Encrypted Postgres database, key derived from an admin password and never persisted in plaintext
• API-key authentication with role-based (admin/user) access control