​

Enroll Voice Clones from Calls

​

Why enroll from a call

• QA a new IVR or agent voice. Check pronunciation, pacing, and tone of your brand voice against a script before it answers a single call.
• Audition agent voices end to end. Pair a clone with the AI agent rollout pipeline: roll the voice behind the same offline comparison, canary, and regression stages that gate prompt and model changes.
• Keep the brand voice consistent across IVR, outbound dialer, and agent surfaces by pointing them all at one trained clone.

​

1. Enrollment-call provenance (source_call_id)

• A clone trained from a recorded call stores that call’s ID. The clone traces back to a specific recording — enough to answer “which call did this voice come from?” during a review.
• A clone trained from an uploaded sample, or designed from a text prompt, has no source call — it stores the sample URL or prompt as its provenance instead.

1. A recorded enrollment call. Record the call with record: true at call creation (see Record a call), then use the call’s ID as call_id. A short, clean enrollment recording — one speaker, minimal background noise, a minute or three of speech — trains a markedly better clone than a noisy hour.
2. Recording consent on that call. The source call must carry an active recording-consent receipt (consent_state='granted', not revoked). Without it, the create call is refused with VOICE_CLONE_CONSENT_REQUIRED — recording consent is the floor for the audio existing at all. Cloning is a separate consent surface on top of it.
3. A voice-owner attestation (below).

curl -X POST https://api.orbit.devotel.io/api/v1/voice/clones \
  -H "X-API-Key: dv_live_sk_..." \
  -H "Content-Type: application/json" \
  -d '{
    "name": "agent-aria",
    "call_id": "call_abc123",
    "voice_owner_relationship": "employee",
    "voice_owner_consent_text": "Ari S. (staff) signed a voice-release on 2026-08-10 covering IVR and agent synthesis use",
    "biometric_consent_acknowledged": true
  }'

const clone = await orbit.voice.clones.create({
  name: 'agent-aria',
  call_id: 'call_abc123',
});

​

Upload a sample instead

curl -X POST https://api.orbit.devotel.io/api/v1/voice/clones/upload \
  -H "X-API-Key: dv_live_sk_..." \
  -H "Content-Type: application/json" \
  -d '{
    "name": "narrator",
    "audio_sample_url": "https://media.example.com/samples/narrator.mp3",
    "voice_owner_relationship": "actor_with_release",
    "voice_owner_consent_evidence_url": "https://media.example.com/consent/narrator-release.pdf",
    "biometric_consent_acknowledged": true
  }'

​

Design a synthetic voice instead

​

2. Watermarking and consent posture

• Recording-consent receipt on the source call — active, not revoked, and no older than one year. Recollect consent before cloning from an old recording.
• A voice-owner attestation you supply at create time. Either a voice_owner_consent_evidence_url (an https:// link to a consent record you store — a signed release, an envelope, an in-app form) or a voice_owner_consent_text (at least 50 characters, stored verbatim), plus:
  • voice_owner_relationship — one of self, employee, actor_with_release, customer_with_written_consent, or other.
  • biometric_consent_acknowledged: true — your explicit acknowledgement of biometric-voiceprint obligations.
  The attestation text travels with the voice to the synthesis provider and is mirrored into your Audit Log, so the legal basis for each clone is one query away.
• Watermarking the output. In the dashboard, open a ready clone and choose Watermark; the dialog marks an audio clip with a provenance credential bound to the exact bytes and to your tenant. Over the API:
  • POST /voice/clones/{id}/watermark — supply audio_base64 (the raw clip) or a pre-computed audio_sha256 + audio_bytes. When raw audio is supplied, the mark is also embedded into the audio itself as a standard content credential any compatible verifier can read; when your AI-disclosure settings enable a spoken preamble, the configured audible disclaimer travels with the mark.
  • POST /voice/clones/watermark/verify — verify a mark, with audio-in-hand binding: a valid mark re-stapled onto different audio reports a mismatch instead of passing.
  Only a clone in status: ready can mark output; a clone still training or one that failed is refused. Whether marking satisfies a specific framework (EU AI Act Art 50, FCC synthetic-voice rules) depends on how you operate the account — the watermark is a control you own; the posture remains yours.

​

3. Provider tracking (provider + provider_voice_id)

• provider — which synthesis provider hosts the trained voice (for example, elevenlabs).
• provider_voice_id — the voice’s identifier at that provider, once training succeeded. If you ever need to cross-reference a voice in the provider’s own console, this is the join key.

​

4. Pricing (price_cents)

• The charged amount is stored on the clone as price_cents (the free clone stores 0) and shown as a price badge next to the clone’s name in the dashboard.
• If the wallet balance is too low, the create call returns 402 with INSUFFICIENT_BALANCE before any work starts.
• If provider training fails after a paid clone was charged, the credits are refunded automatically and the clone is marked failed.

​

5. List, create, and delete

​

6. Status and failure handling

• processing — the sample was accepted and is training at the provider. Most clones finish shortly; the dashboard refreshes the list in place.
• ready — training succeeded, provider_voice_id is set, and the clone can synthesize and watermark output.
• failed — training did not complete: the provider rejected the sample, the sample URL could not be fetched, or the provider call failed. Any charged credits are refunded when this happens. The dashboard shows a failed badge and surfaces the underlying reason as a readable toast — the failure is never silent. Delete the failed clone and re-enroll from a cleaner sample.

​

Errors you will hit and what they mean

​

Next steps

• AI agent rollout pipeline — put a new clone through offline comparison, canary, and regression stages before it answers production calls.
• Audit Log — clone creation, deletion, and watermark events each write an audit entry with the consent basis attached.
• Record a call — capture the enrollment call with record: true so it is clone-eligible.