Speech-to-text that produces searchable, summarised, usable records
Streaming and batch transcription with speaker diarisation, clean-up, summaries and search, built into our own Eventify session transcription product and available for your meetings, calls and content.
Seven stages from microphone to a record people use
A raw transcript is rarely the product. The value is in who said what, cleaned up, summarised, indexed and protected. Each stage below is a place where quality is won or lost.
The transcription pipeline we built for Eventify sessions and reuse for meetings and calls. Stages 5 to 7 are where a transcript becomes something people search and act on.
Built in our own product
Session transcription in Eventify
Eventify, our white-label event platform, transcribes conference sessions and gives attendees summaries and searchable notes after each talk. Conference rooms are a hard environment: distant microphones, applause, several speakers and specialist vocabulary.
Building it taught us where the real work is: capturing clean audio from the AV desk rather than the room, handling speaker changes on panels, feeding session titles and speaker names into the vocabulary, and summarising in a way that respects what was actually said. Those lessons now go into every transcription project we deliver.
Live sessionsPanels and Q&ASummariesAttendee search
Audio quality: microphone placement and the feed you capture from
Accents and speaking styles across your speakers
Domain vocabulary: product names, acronyms, people
Cross-talk and overlapping speakers
Background noise, music and applause
Chunking choices in streaming mode
We measure word error rate on your real audio before and after each change, rather than quoting vendor figures.
How to decide
Streaming versus batch
Concern
Streaming
Batch
Latency
Words within seconds of being spoken
Minutes after the recording ends
Accuracy
Slightly lower; the engine cannot look ahead
Higher; full context, second passes possible
Cost
Higher per minute; always-on infrastructure
Lower; can be scheduled and batched
Diarisation
Harder in real time
Better with the whole recording
Use cases
Live captions, voice agents, real-time note-taking, accessibility
Meeting records, session archives, call analytics, media libraries
Typical design
Live draft for the room, batch pass for the record; the same pipeline, two modes
What we deliver
A transcription service, not a transcript
Capture and recognition
Integration with your AV, telephony, meeting tools or app; engine selection and custom vocabulary; streaming and batch modes.
Understanding layer
Summaries, action items, topics and decisions extracted with an LLM, graded against an eval set so they stay faithful. The engine behind meeting intelligence.
Storage, search and control
Semantic search over transcripts, retention and redaction rules, per-user access, and audit of who listened to or read what.
Sample and measure
Real recordings, a ground-truth transcript for a subset, baseline word error rate.
Build the pipeline
Capture, recognition, diarisation, clean-up, summaries, storage. Evals for the summaries, not just the words.
Tune
Vocabulary, chunking, noise handling, speaker handling. Measured against the same sample.
Launch and watch
Rollout, monitoring of latency, cost and error rate, feedback loop from corrections.
What goes wrong
Summaries that invent decisions nobody made, transcripts that attribute words to the wrong speaker, and audio retained forever because nobody set a policy. All three are design failures, and all three are in our week-one checklist.
Streaming when someone needs the words as they are spoken: live captions, voice agents, real-time notes. Batch when the recording already exists and accuracy or cost matters more than immediacy: meetings, sessions, call archives. Many systems do both: a live draft, then a cleaner batch pass.
How accurate is the transcription?
It depends on audio quality, accents, vocabulary and cross-talk far more than on the engine. We measure word error rate on a sample of your real audio before quoting anything, then improve it with custom vocabulary, better capture and post-processing. See the glossary entry for the terms.
Have you shipped this in production?
Yes. Session speech-to-text is a feature of our own event platform, Eventify, where talks are transcribed and summarised for attendees. The case study covers the pipeline and what we learned about noisy rooms and multiple speakers.
Can transcripts stay inside our environment?
Yes. We can run open-weight speech models on your infrastructure or in a region you choose, and we design retention, redaction and access rules with our AI data privacy practice. Audio is often more sensitive than text; we treat it that way.
How long does a transcription project take?
A batch pipeline with summaries and search typically takes four to eight weeks. Streaming with live captions or a voice agent front end takes six to ten, mostly on latency tuning and the front-end integration.
Ready to make AI work in production?
Tell us about your systems, your data and your constraints. We reply within one business day with a straight answer and a suggested next step.