Deployment & integration
Operations & IT
You are answerable for where the audio goes, and a public endpoint is not an answer.
Enterprise
On-premises, self-hosted enterprise — inside your walls
On-premises deployment, a speech model trained on your own vocabulary, a full API, and a team you can call — for the volumes and the security requirements a self-serve plan was never meant to carry.
- On-premises
- or your cloud, or ours
- Zero egress
- audio never leaves the estate
- Your own model
- trained on your vocabulary
- Named support
- commitments in the contract
If it is spoken at your company, this page is for you.
Four departments, four kinds of recording, one requirement — it has to stay inside your walls.
Calls & quality review
Contact centres
Thousands of calls a week have to be searchable, scored and kept for the auditor.
Hearings & minutes
Legal & compliance
The record has to be verbatim, attributable, and retained on a schedule you set.
Subtitles at volume
Media & comms
A back catalogue has to ship in every market, on a house spec, without a bottleneck.
Three ways to run it. The same build in all three.
Pick the one your security review can live with — moving between them later is a configuration change, not a migration.
On-premises
Regulated, or air-gapped
- Your own hardware, your data centre
- No egress — no route to the public internet
- You hold every key and every log
- Longest to stand up; nothing leaves the building
Private cloud
Cloud-first with residency rules
- Your cloud tenancy, your region
- Traffic stays inside your VPC
- You own the account; we deploy into it
- Standard infrastructure, standard procurement
Managed
Fast start, standard risk
- Our Hong Kong region
- TLS in transit, encrypted at rest
- You administer users and retention
- Running the same week you sign
Whichever you pick
Data processing agreementNamed sub-processor listSecurity questionnaire on filePenetration test summaryAudit log exportSSO — SAML or OIDCSCIM provisioningCustom retention windows
Nine reasons this survives procurement.
Not a feature list — the things a security review, a finance team and an auditor each ask about.
A model trained on your words
- Your terms, not a guessProduct names, drug names, case numbers — learned from data you supply.
- Your industry's speechTrained against the accents, jargon and call quality you actually record.
- Retrained as you goCorrections feed the next version instead of being made again every week.
- Measured before it shipsYou see the accuracy on your own held-back recordings, not on a benchmark.
A real API, not an export button
- REST for the whole productSubmit, poll, fetch and delete — everything the interface does, over HTTP.
- Webhooks when it finishesNothing to poll on a timer; the job tells your system when it is done.
- Documented and sandboxedAn OpenAPI spec and a sandbox key, so integration starts before the contract.
- Keys per environmentStaging and production are separate tokens with separate quotas.
It can run inside your walls
- On your own hardwareThe full pipeline deploys on infrastructure you control and audit.
- No egress at allAudio never crosses to the public internet, so residency stops being a question.
- Your compliance regimeRetention, access and logging configured to whatever you are held to.
- Same build everywhereMoving between on-premises and managed later is configuration, not migration.
Workflows shaped to your process
- It starts itselfA watch folder, a calendar or a webhook — no one has to remember to upload.
- Your rubric, your rulesScoring and routing follow the criteria your team already reviews against.
- Ends where the work isResults land in your warehouse, CRM or case system, not in another inbox.
- Built with you, onceWe design the pipeline during onboarding rather than handing you a toolkit.
Support with a name on it
- A named account managerOne person who knows your deployment, not whoever picks up the ticket.
- A shared channelYour engineers and ours in the same room, in Slack or Teams.
- Training for your teamOnboarding sessions run for the people who will actually use it.
- Commitments in writingResponse and uptime targets sit in the contract, not on a status page.
Backed up, and provably restorable
- Automatic, every six hoursNothing depends on somebody remembering to run an export.
- Held in three placesSeparate locations, so losing one of them costs you nothing.
- Restore inside four hoursA target we rehearse against, not a promise made after the incident.
- You set the retentionAnd deletion is real deletion, logged where your auditor can see it.
Cantonese, written how you publish
- Spoken form, then written口語 transcribed as said, then converted to the 書面語 you actually publish.
- Both versions keptAligned turn for turn, so a reviewer can check the polish against the record.
- The case we build againstNot an afterthought bolted onto an English model that guesses at the rest.
- Your house styleTraditional or simplified, and the terminology your editors already enforce.
Built for the hard recordings
- Phone-line audio8kHz call-centre recordings, not just a clean microphone in a quiet room.
- Noise and crosstalkSite visits, ceiling mics and people talking over each other.
- Proven on yours, not oursWe run your worst recordings during evaluation and show you the numbers.
- Where the savings areThese are the files a human currently has to redo from scratch.
It knows who said which line
- Speakers kept apartEach turn attributed, so minutes and quality scores can be defended.
- Name them onceRename a speaker and it follows into every turn and every export.
- Language detected liveIncluding a switch mid-sentence, which is the norm in a Hong Kong meeting.
- Timestamps that holdEvery line can be found again in the audio when somebody disputes it.
And the rest of the checklist.
None of these is the reason to pick us, but every one of them comes up somewhere between the demo and the signature.
SSO with SAML or OIDCSCIM user provisioningRole-based permissionsAudit log exportWorkspace-wide glossaryUsage reporting per teamInvoiced annual billingCustom retention windowsIP allowlistingData processing agreementSecurity questionnaire on fileNamed sub-processor listVolume pricingMulti-workspace administration
If something here is a blocker for you, it is worth one call to find out early.
Book a demoQuestions that come backfrom security and procurement
Yes. The full pipeline — ingest, transcription, diarisation, storage — deploys on hardware you control, with no route to the public internet. The same build also runs in your private cloud tenancy or in our managed region, so if you start managed and move on-premises later, that is a configuration change rather than a migration.
No. Nothing you upload is used to train shared models. A model trained for you is trained only on the data you supply for that purpose, and it is yours — it is not folded back into the general model other customers use.
A set of recordings that look like your real ones, and your terminology — product names, drug names, case references. We hold back a portion of what you send, train on the rest, and report accuracy on the held-back set so the number you see was measured on your audio rather than on a public benchmark. Corrections your team makes later feed the next retrain.
Everything the interface does: submit, poll, fetch, and delete, over REST, with a webhook when a job completes so you are not polling on a timer. There is an OpenAPI spec and a sandbox key, and staging and production hold separate tokens with separate quotas — so integration can start before the contract is signed.
You set the retention window; when it expires the data is purged rather than flagged. Deletions are written to an audit log you can export, which is usually what an auditor is asking for when they ask this question.
You have a named account manager and a shared channel in Slack or Teams with the engineers who built the deployment, not a ticket queue. Response and uptime commitments sit in the contract. Backups run automatically every six hours to three separate locations, and we rehearse restores against a four-hour target.
On committed volume rather than per seat, invoiced annually. On-premises deployment, a custom model, and the support commitments are priced separately from usage, because most of that cost is one-off. The demo is the fastest way to get a number that means anything.
Questions that come backfrom security and procurement
Yes. The full pipeline — ingest, transcription, diarisation, storage — deploys on hardware you control, with no route to the public internet. The same build also runs in your private cloud tenancy or in our managed region, so if you start managed and move on-premises later, that is a configuration change rather than a migration.
No. Nothing you upload is used to train shared models. A model trained for you is trained only on the data you supply for that purpose, and it is yours — it is not folded back into the general model other customers use.
A set of recordings that look like your real ones, and your terminology — product names, drug names, case references. We hold back a portion of what you send, train on the rest, and report accuracy on the held-back set so the number you see was measured on your audio rather than on a public benchmark. Corrections your team makes later feed the next retrain.
Everything the interface does: submit, poll, fetch, and delete, over REST, with a webhook when a job completes so you are not polling on a timer. There is an OpenAPI spec and a sandbox key, and staging and production hold separate tokens with separate quotas — so integration can start before the contract is signed.
You set the retention window; when it expires the data is purged rather than flagged. Deletions are written to an audit log you can export, which is usually what an auditor is asking for when they ask this question.
You have a named account manager and a shared channel in Slack or Teams with the engineers who built the deployment, not a ticket queue. Response and uptime commitments sit in the contract. Backups run automatically every six hours to three separate locations, and we rehearse restores against a four-hour target.
On committed volume rather than per seat, invoiced annually. On-premises deployment, a custom model, and the support commitments are priced separately from usage, because most of that cost is one-off. The demo is the fastest way to get a number that means anything.