Real-time transcription
Live captioning and streaming transcripts with sub-second latency for calls, meetings and accessibility use.
Your organisation already records an enormous amount of speech — calls, meetings, site visits, consultations. Almost none of it is searchable. Transcription turns that into text; the useful part is everything you build on top of the text.
A raw transcript is rarely what anyone wants. What they want is the decision that was made, the commitment given, the risk mentioned, or a record in the CRM. We build the layer that produces those.
Model selection and tuning for your audio conditions — phone lines, meeting rooms, noisy sites — rather than one generic setting.
Who said what, so a transcript reads as a conversation and can be attributed.
Product names, drug names, part numbers and internal jargon recognised correctly instead of phonetically guessed.
Summaries, action items, sentiment and structured fields produced from the transcript and written where they belong.
Live captioning and streaming transcripts with sub-second latency for calls, meetings and accessibility use.
Historical recordings processed in bulk to make an existing archive searchable for the first time.
Summaries, decisions and owned action items extracted automatically and distributed to attendees.
Talk ratio, topic detection, objection tracking and compliance phrase checking across contact-centre volume.
Transcription across languages, including conversations that switch language mid-sentence.
Automatic removal of card numbers, identifiers and personal data from transcripts and audio before storage.
Real-time and batch have genuinely different designs, and pretending otherwise causes most of the problems we get called in to fix. We build the one your use case needs — or both, sharing the same post-processing.
WebSocket audio in, partial results out, with buffering that balances latency against accuracy.
Queued processing with chunking and merge, sized for cost rather than speed.
Punctuation, casing, vocabulary correction, diarisation and redaction applied consistently to both paths.
Word error rate tracked on a labelled sample of your own audio, so a model change is a measured decision.
Capability in isolation changes nothing. These are the systems this normally has to reach in order to be useful.
Recording pipelines and live audio from the platform you already use.
Capture from the tools your teams actually meet in, with attendee context preserved.
Structured outcomes written to the right record rather than dumped into a notes field.
Transcripts indexed for retrieval so spoken knowledge becomes answerable — see our RAG work.
A sales call from hang-up to a fully updated CRM record, with no manual note-taking.
The same six stages apply to every AI build we run. The evaluation set is created before development starts — without it there is no way to tell an improvement from a regression.
Is a model the right instrument here? A deterministic rule is cheaper, faster and easier to audit when the problem allows one. We say so when it does.
Sources identified, access agreed, content parsed and indexed. Data quality sets the ceiling on everything after this point.
Real examples with known-good outcomes, collected before any development. This is what makes later decisions measurable.
A working slice scored against the evaluation set — accuracy, latency and cost per request all measured rather than estimated.
Wired into the real system with permissions, rate limits, spend caps, fallbacks and human approval wherever the action is irreversible.
Quality, cost and drift tracked after launch, with the evaluation set re-run whenever a prompt, model or data source changes.
These are not add-ons quoted separately. They are the conditions under which a system like this is allowed near production data.
Recording consent handled explicitly, including per-jurisdiction rules where you operate across borders.
Sensitive values removed from both transcript and audio at ingest, not on retrieval.
Open-weight ASR models deployed in your own environment where recordings must not leave it.
Automatic deletion schedules enforced by the system rather than by a reminder in someone's calendar.
Consultation notes drafted from the conversation so the clinician can look at the patient rather than the keyboard — reviewed and signed before filing.
Review every call for required disclosures and escalation triggers instead of sampling a small percentage.
Attributed, timestamped transcripts of long recordings, searchable across an entire matter.
Call outcomes, next steps and objections written to the CRM record automatically after the call ends.
Spoken observations captured on site, transcribed and filed against the correct project and location.
Make years of interviews and broadcasts findable by phrase, speaker and topic.
Accuracy depends far more on your audio conditions than on the model. Clean meeting audio typically reaches the mid-to-high nineties in word accuracy; compressed phone audio with cross-talk is meaningfully lower. We measure word error rate on a labelled sample of your own recordings during discovery, so you get a real number rather than a vendor benchmark.
Yes. Domain vocabulary — product names, part numbers, clinical or legal terms — is supplied as a custom vocabulary or handled with a correction pass over the transcript. This is normally the single biggest accuracy improvement available, and it is cheap to do.
Yes. Open-weight models such as Whisper run entirely on your own hardware or cloud tenancy, so audio never reaches a third party. For regulated environments this is usually the right choice, and we size the compute during discovery.
Yes, through diarisation, which separates and labels speakers. Accuracy is good with distinct voices and separate microphones, and degrades with heavy cross-talk on a single channel. Where attribution is critical we recommend capturing separate channels, which makes the problem far easier.
Send us a sample of your real audio. We will measure achievable accuracy and show you what could be extracted from it.