Records live or accepts uploaded audio of any meeting, performs ASR via Gemini 2.5 Flash, runs speaker diarization, then extracts a structured minutes document. Attendees with talk-time share, topic-grouped summary, action items with assignee + due, decisions with rationale, risks with severity — every line carries an audio anchor (▶ MM:SS) that scrubs the recording to the exact moment. Action items auto-promote to the Tasks tracker.
Construction meetings produce decisions that bind. The OAC meeting confirms the chiller substitution. The pre-installation meeting locks the framing sequence. The owner walkthrough records 14 items that became RFIs. None of it survives unless someone writes minutes — which usually means someone listening less than they should, transcribing rough notes, and emailing a summary nobody reads twice.
Voice Synthesis replaces that workflow. Record the meeting (mobile or desk), or upload audio after. The engine transcribes with speaker diarization (knows who said what), then composes structured minutes: attendees with talk-time share, topic-grouped narrative, action items with assignee and due date, decisions with rationale, risks with severity. Every output line carries a ▶ MM:SS audio anchor — click it, the recording scrubs to the exact moment.
The engine handles construction-domain audio — site noise, multiple speakers, technical vocabulary — better than generic transcription tools. Built for the meetings construction actually runs:
Live recording (mobile or web) or uploaded audio file. Context inputs improve accuracy but aren't required.
Live capture or upload. Common formats accepted: M4A, MP3, WAV, OGG. Up to 4 hours per file. Mobile mic, desk mic, or conference room speakerphone all work.
Project name + attendee list + recent meetings. The engine uses this to tune diarization (knows which speaker is the architect vs the GC PM) and to disambiguate terms (knows "Carrier" is the chiller manufacturer in this project).
Open RFI / submittal / procurement state lets the engine cross-link mentions in the meeting to existing trackers. "Did the chiller submittal come back?" auto-links to SUB-0247 instead of becoming a vague reference.
The minutes render with audio anchors per line. Click ▶ on any output and the recording scrubs to that moment. Below: a sample synthesis from a 47-minute OAC meeting.
Two weeks later when the sub disputes "you said we'd resequence the framing," the PM clicks ▶ 11:30 and plays the actual decision back. Six months later when the owner asks why the chiller is Carrier not Trane, the audio anchor surfaces the OAC decision. Source-of-truth that doesn't require trusting anyone's memory.
Voice Synthesis is built to be the multi-tracker entry point. Field meetings produce actions and decisions for several trackers at once — the engine extracts and routes each into the right place.
Gemini 2.5 Flash handles typical construction trailer / conference room audio at >95% word accuracy. Active site audio (generators, hammer drills) drops to ~85% on the noise side — meeting content typically stays clear because mics are pointed at speakers. Domain vocabulary (RTU, AHU, VAV, AIA G702) is correctly transcribed by default; the engine doesn't need a glossary.
If you supply an attendee list with project context, the engine matches voice profiles across recordings (the architect on this project always says certain phrases, has a certain cadence). First meeting: speakers labeled as Speaker 1 / 2 / 3 with the option to tag once. Subsequent meetings auto-recognize. Accuracy improves with project meeting count.
Recordings are stored encrypted at rest, scoped to the project's authorized users. PII detection flags personal data for review before exporting. Off-the-record portions can be marked at recording time (pause + resume) so they don't synthesize into the minutes. Audio retention defaults to 18 months; configurable to 1–7 years per project.
Yes. Upload any audio file (recorded on a phone voice memo, dictated post-meeting, etc.) and the engine processes it the same way. Live mode just saves you the step of uploading after.
Action items with a clear assignee + due date auto-promote to the Tasks tracker (source = meeting). Ambiguous actions ("someone needs to check the panel") stay in the meeting summary for PM review. The PM can promote any line to Tasks with one tap. Default behavior favors auto-promotion to avoid actions getting lost.
Bring a 30-minute OAC or pre-installation recording to a 15-minute call. We'll process it on screen and walk through every action, decision, and RFI extracted. You keep the minutes and the audio.
Start free — 14 days, no card →