Identity & access
Role-based access, least privilege and MFA for administrative access. Access is reviewed when a person changes role and removed when they leave.
Trust / Security & responsible voice
Singapore · Contract-led
A voice agent is a promise you can hear.
Security, consent and data handling are part of the product, not a page added before an audit. This is how the stack is controlled, and what we will not do with it.
Security contact
Security contact form
Acknowledged within one business day.
Acknowledged within one business day.
Template policy: replace these target controls and operational claims with your actual controls before publishing.
Privacy
Singapore data protection
TLS / SRTP
In transit, including media
0
Third-party ad trackers on this site
Contract
Retention and deletion defined per client
On this page
Security is a control, not a certificate.
We run a small number of systems with a high consequence of failure. The controls are written down, tested and reviewed. A badge is a by-product, not the objective.
Encryption
TLS 1.2+ for signalling and APIs, SRTP for media, AES-256 at rest. Keys are managed separately from the data they protect.
Isolation
Per-tenant runtime isolation. Dedicated and sovereign deployments get separate capacity, credentials and retention rules.
Change management
Code review, automated tests, canary releases and a rollback. A model change follows the same gate as a code change.
Monitoring
Centralised logs, per-stage latency, alerting and a 24×7 on-call rotation. We alert on caller impact, not on CPU graphs.
Recovery
Encrypted backups, restore tests and a business continuity plan. The plan is exercised, not filed.
On attestations: we do not display a certificate or badge we cannot produce on request. Current security attestations and scope are disclosed during procurement.
On incidents: every P1 gets an incident lead, a timeline and a written review. Notification timelines are agreed in the contract and met.
Data follows the contract.
We process two kinds of data: our own business contacts, and the call data we process on behalf of a client. The rules are different, and the contract says which is which.
Our own business contacts
When you contact us, request portal access or submit a ticket, we collect the details you provide and basic technical logs. We use them to answer you, operate the portal and keep the service secure. We do not sell them, and we do not add you to a marketing list without a separate opt-in.
Client call data
For a client deployment, Aldaria acts as a processor. Call audio, transcripts, summaries and tool payloads are processed under the client's instructions. Recording is off by default unless the flow and the contract enable it. PII redaction can run before storage.
Residency, retention, deletion and subprocessors are written into the deployment record. Where data moves between Singapore · Panama · Hong Kong · China, the transfer mechanism and the permitted purpose are agreed in advance.
Typical retention
| Data | Typical retention | Basis |
|---|---|---|
| Contact and ticket data | 24 months after opening | Business relationship |
| Portal account | Active + 12 months | Access and support |
| Server logs | 90 days | Security and operations |
| Call audio / transcripts | Per client contract | Processor instructions |
| Consent records | Per regulation and contract | Legal obligation |
No training on customer data by default. Client recordings and transcripts are not pooled into a shared model. Any use for improvement is explicit, scoped and instructed in writing. Read the privacy policy.
03 / Responsible voice
Consent is an interface.
A voice agent speaks with your brand and touches real people. Disclosure, consent and escalation are designed into the call, not appended to a policy page.
- Emergency services or emergency dispatch.
- Political persuasion or campaign calling.
- Unsolicited robocalls or purchased lists.
- Impersonation of a real person without consent.
- Voice clones without documented, revocable consent.
- Harassment, coercion or deceptive debt collection.
Disclosure. The agent identifies itself and its organisation at the start of the call, and says when the call is recorded. The disclosure is part of the first turn, not a menu the caller can skip.
Consent. Outbound campaigns require a lawful basis, a suppression check and a calling window. A clone requires written, specific and revocable consent from the person whose voice is used.
Escalation. A caller can ask for a human at any point. The agent also escalates on failed authentication, policy boundaries and low confidence. The handoff carries context.
Audit. Consent records, suppression lists and call metadata are retained to evidence the rules were followed. Recordings follow the client's retention instructions.
Models are governed like infrastructure.
A voice model is a production system with a version, a baseline and a rollback. It is not a prompt update pushed on a Friday.
Evaluation
Replayable calls, a regression suite, WER and MOS tracking, and listener review. The scorecard is agreed before launch.
Accent & language
Test coverage across the accents, languages and code-switching patterns in the deployment. Errors are reported by group, not averaged away.
Human oversight
Policy boundaries, escalation thresholds and a human review queue. Automation does not get the final word on an adverse action.
Provenance
Model versions, training data categories and changes are recorded. A voice or model can be traced to the release that produced it.
Red teaming
Prompt injection, social engineering, tool abuse and data exfiltration are tested before launch and after material changes.
Incident reporting
Model or voice incidents follow the security process. Clients are told what happened, what changed and what was re-tested.
The obligations we can name.
We do not claim every framework. We name the obligations that apply to our business and our clients, and we show the evidence in procurement.
| Framework | Status | What it means |
|---|---|---|
| Singapore PDPA and applicable data protection law | Applies | Personal data is handled under the applicable privacy law, with a privacy contact route. |
| International privacy laws | Where applicable | Controller for business contacts; processor for client call data, with a DPA. |
| Security attestations | As applicable | Scope and current status are disclosed during procurement. |
| Telecom rules | Applies | Enterprise voice only. Emergency services are out of scope. |
| Client contracts | Contracted | SLAs, retention, subprocessors and deletion terms are agreed in writing. |
We are not a carrier of last resort. Aldaria provides enterprise voice infrastructure and managed agents. We do not operate emergency services, and we will decline a deployment that requires them.
Report a vulnerability.
If you believe you have found a security issue in our site, portal or platform, tell us before you tell anyone else. We will investigate and fix it.
- Contact: use the security contact form. Ask for the current encrypted-reporting key in your first message.
- Acknowledgement: within one business day. We will give you a named human and a timeline.
- Safe harbour: we will not pursue legal action for good-faith research that respects privacy and does not degrade service.
- Scope: production systems only. Do not test against a client's data, a live call or a third-party service.
- Bounty: we do not currently run a paid bounty. We will credit a fix if you want to be named.
Privacy questions: use the privacy contact form. The full policy is on the privacy page.
Start a conversation
Bring us the call.
Tell us what the call has to do. We will tell you what it takes to run it.