Trust, safety and how we protect voices
This page describes the measures we actually operate, so that a voice owner, a customer or a reviewer can check them. Everything in the first list runs in production today; everything in the last section is not yet built, and is written in the future tense on purpose.
1. Controls in operation today
Each item below is enforced by the service itself, not only by a policy document. Where we say a request is refused, it is refused on the API as well as in the interface.
1.1. A cloned voice cannot be created without an explicit attestation
Registering a cloned voice requires two separate statements from the caller: that they hold the right to use the voice, and which right they claim — their own voice, someone else's voice with documented authorisation, or a voice their organisation is licensed to use. A bare confirmation is refused by the server with a machine-readable error; the interface cannot be bypassed by calling the endpoint directly.
The same requirement applies to cloning from a recording uploaded with each request, which is a separate code path and used to be the gap.
Evidence: POST /v1/voices and POST /tts/clone both call the consent validator, which raises consent_type_required when the type is missing.
1.2. What the consent record contains
For each cloned voice we store: the account, the voice id, the consent type, the version of the policy the person accepted, the moment of acceptance, an optional reference to the written authorisation they hold, and the network metadata of the request. We deliberately do not ask for identity documents or biometric templates.
The policy version is bumped whenever the wording of the consent policy changes, so a voice can never claim to have accepted text it never saw. An account can read its own record back; the network metadata is not returned to clients.
1.3. Traceability of generated audio
Every generation from a cloned voice writes a traceability row: a generation identifier, the account, the voice, the moment, the route and the length of the input. The identifier is returned to the caller in the X-OV-Generation-Id header, so a person who finds a clip can quote a reference that resolves to exactly one account and one voice.
The row never contains the generated text or the audio, and it is not public. It is readable only by the operator through the moderation queue, which is what turns "somewhere on your platform" into an actionable complaint.
Evidence: backend/voice_safety.py, collection voice_generation_audit; lookup in GET /admin/moderation/reports/{ticket}.
1.4. Disabling a voice takes effect everywhere
A voice can be disabled during a review, or revoked when its owner withdraws consent or a licence ends. Either state is enforced where the voice is resolved, which is the single point every synthesis route passes through — the studio, the native text-to-speech API, the OpenAI-compatible endpoint and the MCP connector.
So a disabled voice stops producing audio on all of them, and the caller gets an explicit error naming the state, not a confusing timeout or a silent fallback to a different voice. Deleting a voice removes its reference recording, its derived clip and its cached prompt, and a deleted voice cannot be used to generate anything.
1.5. Account-level enforcement
Voice cloning can be suspended on a single account while an investigation is open. The suspension is narrow on purpose: it blocks new registrations and every use of a cloned voice, while leaving the account able to sign in, read its library and export its own data. A disputed voice does not quietly take a customer's whole workspace offline.
1.6. A reporting channel with a real ticket
Anyone — with or without an account — can file a report through the abuse report form. Each report is stored as a durable ticket with a reference, and moves through a recorded lifecycle: submitted, initial review, risk assessment, evidence requested, actioned or rejected, closed.
If our storage is unavailable the form says so and points to email rather than showing a success message for a report we did not keep. Deliberately false reports are recorded too and are a breach of our acceptable use policy.
1.7. Volume and cost limits
Voice registration is rate limited per account, uploads are size limited, and the number of cloned voices an account may hold is capped by its plan. Report intake is rate limited per connection and de-duplicated, so the queue cannot be flooded from one address. These are cost and abuse brakes, not identity checks.
1.8. Your recordings are not used to train a model
SaydiVoice does not train, fine-tune or otherwise improve a voice model using your recordings or your generated audio. A cloned voice is used as a conditioning reference to render the audio you asked for; it does not become part of a model. Transcripts are produced by a speech-recognition processor purely to align that reference, and the reference is deleted when you delete the voice.
2. What these controls do not do
Being precise about the limits is part of the control being trustworthy.
- A consent attestation is a statement, not proof of identity. We record who claimed what, and we can hold them to it; we cannot, by ourselves, verify that the person uploading a recording is the person it sounds like.
- We do not scan every generation for policy compliance before it is produced. We act on reports, on risk signals and on the evidence in the traceability log.
- We do not embed an inaudible watermark in generated audio. Traceability works through the generation record and the account, not through the waveform, so it does not survive a re-recording and it is not a forensic identifier.
- A takedown is not a legal determination. It is a platform decision we take to stop harm while a claim is resolved; it does not decide who owns the voice.
3. How a report moves through our team
Report submitted → ticket recorded with a reference → initial review by a person → risk assessment against the traceability record → evidence request to the account holder, and to the reporter where needed → decision → enforcement applied in the service → notification with reasons → appeal or re-review → closed.
Every step is written to the ticket's history with a timestamp and the reason, so a decision can be reviewed later without relying on anyone's memory. The timescales we commit to are published on the report page.
4. Data protection for voice samples and reports
Voice recordings, generated audio, consent records, traceability rows and abuse reports are personal data and are handled as described in our Privacy Policy and, for business customers, the Data Processing Appendix. In summary:
- Access is restricted to the account that owns the data and, for investigation, to the operator's moderation tools.
- Voice data is not published: a cloned voice is reachable only through the owner's account, and the public catalogue is a separate resource.
- Deleting a voice removes the reference recording, the derived clip and the cached prompt. Deleting an account removes its voices; traceability rows and report records are retained for the period needed to handle appeals and to meet legal and accounting duties, then deleted or anonymised.
- We disclose information to a competent authority where the law requires it, and we keep a record of such requests.
5. What we are still building
These are planned and are not active. We are listing them so that nobody reads a commitment into the sections above that is not there yet.
- Voice verification at registration — a read-aloud challenge (a "voice CAPTCHA") with liveness and speaker-similarity scoring, to raise the barrier for cloning a voice from a recording obtained elsewhere. Speaker similarity is a risk signal, not proof of identity, and it will be described as such when it ships.
- Machine-readable marking of generated audio, to the extent the implementing rules under Vietnam's AI Law require it of a provider, together with labelling guidance for deployers.
- Automated risk scoring and behaviour detection to route suspicious registration and usage patterns into the review queue sooner than a report would.
- Automated retention enforcement for voice artefacts, so a retention period is applied by the system rather than by an operator remembering to.
6. Contact
Safety reports: the report form. Everything else, including data-protection requests and law-enforcement enquiries: contact@saydi.ai.
This page describes the service honestly at the date shown at the top. When a planned control ships, it moves into section 1 with the date it took effect.
Questions people ask us
Can I clone someone else's voice if I have their permission?
Yes, if the permission is real, specific to what you are doing, and you can evidence it. Register the voice with the "someone else's voice with documented authorisation" consent type and keep the written authorisation. You must not exceed the scope that was granted, and the person can withdraw it — at which point the voice is revoked and stops rendering.
Does SaydiVoice verify that the person uploading a voice is really that person?
Not today. We require and record an explicit attestation of rights, and we act on reports with the evidence in our traceability log — but automated voice verification is on the planned list, not in production. We would rather say that clearly than claim a check that does not exist.
How quickly will you act on a report?
A person reads it and replies within 24 working hours. If the risk is significant the voice can be disabled before that review completes. We do not commit to a fixed removal deadline, because the honest answer depends on whether the account holder produces a valid authorisation.
If I delete my cloned voice, can it still be used?
No. Deleting a voice removes the stored reference recording, the derived clip and the cached prompt, and the voice id stops resolving immediately on every surface, including other server instances.
Do you use my recordings to train models?
No. Your recordings are used as a conditioning reference for the audio you request, and are not used to train or fine-tune a model.
Legal sources cited
The instruments below are the legal bases cited in this document. The links point to the Government legal portal and the signed original.
- Law on Artificial Intelligence No. 134/2025/QH15 — passed by the National Assembly on 10 December 2025, in force from 1 March 2026
Government legal portal · Signed original (PDF) - Law on Personal Data Protection No. 91/2025/QH15 — passed by the National Assembly on 26 June 2025, in force from 1 January 2026
Government legal portal - Decree 356/2025/ND-CP detailing the Law on Personal Data Protection — issued by the Government on 31 December 2025, in force from 1 January 2026 and replacing Decree 13/2023/ND-CP
Government legal portal
See the provision mapping and the full text of every cited provision
See also the voice consent policy, the AI voice safety and acceptable use policy and the report page.