Voice agents for legal transcription and proceedings

AssemblyAI's Universal-3.6 Pro Realtime is the leading speech-to-text API for deposition and legal transcription — a 307 ms median time to final transcript (Pipecat open benchmark), native speaker diarization for counsel/witness/judge separation, and dynamic keyterm prompting (up to 100 terms per session) for statutes, party names, and matter-specific vocabulary. Inline PII redaction masks SSNs, dates of birth, and addresses before transcripts reach your case management system. Universal-3.6 Pro Realtime runs at $0.45/hr with unlimited concurrency. For voice-driven legal intake (new-client calls, scheduling, basic Q&A), pair it with an LLM through LLM Gateway and the Text-to-Speech API so the agent can talk back — streaming STT, LLM Gateway, and TTS on one API key and one bill.

Stream deposition or hearing audio into Universal-3.6 Pro Realtime via WebSocket. Enable speaker_labels=true to attribute each turn to a speaker (the model returns a `speaker_label` field per Turn event). Add case names, party names, exhibit numbers, and matter-specific terms as keyterms_prompt (up to 100 per session, 50 characters each) so the model recognizes them accurately. Enable redact_pii=true with redact_pii_policies for the categories you need to mask (SSNs, dates of birth, account numbers). Pipe finalized turns with timestamps into your case management or discovery workflow.

Yes. Universal-3.6 Pro Realtime includes native speaker diarization — enable speaker_labels=true and each finalized Turn event includes a `speaker_label` field (e.g. "A", "B", "C") attributing the turn to a distinct voice. Speaker accuracy improves over the course of a session as the model accumulates embedding context, which is ideal for long depositions and multi-day hearings. For multichannel captures where each speaker has their own microphone, run separate WebSocket connections per channel for cleaner separation than diarization alone.

AssemblyAI's streaming PII redaction masks personally identifiable information on finalized turns before they leave the API. Set redact_pii=true on the connection and pass redact_pii_policies to scope categories — SSNs, dates of birth, phone numbers, email addresses, account numbers, credit card numbers, addresses, person names, and more. Use redact_pii_sub="entity_name" to replace with category tags like [SSN] or [DATE_OF_BIRTH], or redact_pii_sub="hash" for hash masking. For workflows that need redacted audio files as well, use AssemblyAI's pre-recorded PII redaction with redact_pii_audio=true.

AssemblyAI's Universal-3.6 Pro Realtime integrates with any case management or discovery platform that exposes a webhook or API — Clio, MyCase, Practice Panther, LexisNexis, Westlaw, Relativity, Everlaw, and custom builds. Pipe finalized transcripts (with speaker_label and word-level timestamps) into your platform via REST API or webhook. For voice-driven intake and scheduling, build the agent in LiveKit, Pipecat, or your own orchestrator with Universal-3.6 Pro Realtime, an LLM via LLM Gateway, and the Text-to-Speech API — register lookup_client, create_matter, or schedule_consultation as LLM tools and the agent invokes them mid-call to push data directly to your case management backend.

AssemblyAI's Universal-3.6 Pro Realtime has the lowest word error rate on the Coval STT leaderboard (2.2%, 7-day window). Most jurisdictions still require a certified court reporter to attest to the official transcript — AI transcription typically supplements (not replaces) certified records. The most common pattern: use AssemblyAI for real-time draft transcripts during proceedings, then have a certified court reporter finalize and attest the official version. AssemblyAI is SOC 2 Type 2 certified, with PCI DSS v4.0 and ISO 27001:2022 covering the underlying infrastructure for regulated legal workloads.