Skip to content
Vulpine
Esc
↑↓navigate↵open⌘Jpreview
On this page

Voice profiles

Draft: the manifest a Voice Profile candidate carries before any synthesis (consent record, the four sha256-hashed inputs, a profile id derived from persona, seed and hashes, operator-owned WER and naturalness budgets) and the codes the checker tools/speech/voice_foundry.py refuses it with.

Living doc, checked against the code 2026-11-14

voice-profile-manifest.v1 — the consent and provenance a Voice Profile candidate must carry

  • Status: draft. The checker is built. Nothing in the gateway reads the manifest. The gateway’s Voice Profile is a name (VoiceProfile {id, label, description} in crates/protocol/src/lib.rs) with no provenance field.
  • Checker: tools/speech/voice_foundry.py (check, derive-id).

A Voice Profile candidate is a voice made from reference material plus a model, a config, a prompt and a seed. Its manifest must pass this check before any synthesis, evaluation or admission uses the candidate. The manifest is a JSON file next to its artifacts.

The manifest

{
  "schema": "voice-profile-manifest.v1",
  "profile_id": "vp1-<24 lowercase hex>",
  "profile": {"persona": "fox9", "label": "free text", "uses": ["evaluation"]},
  "consent": {
    "record_id": "consent-0001",
    "subject": "a pseudonym, never a real name in git",
    "granted_on": "2026-10-01",
    "expires_on": "2027-10-01",
    "scope": ["evaluation", "audition"],
    "revoked_on": null
  },
  "inputs": {
    "seed": 7,
    "artifacts": [
      {"role": "source", "path": "source.wav", "sha256": "<64 lowercase hex>"},
      {"role": "model", "path": "model.json", "sha256": "..."},
      {"role": "config", "path": "config.json", "sha256": "..."},
      {"role": "prompt", "path": "prompt.txt", "sha256": "..."}
    ]
  },
  "budgets": {"set_by": null, "wer_regression_max": null, "naturalness_regression_max": null}
}
  • uses and scope use a closed vocabulary: evaluation (offline scoring), audition (a blind A/B for the operator) and persona_tts (a Persona speaks with the voice). Any other word is refused.
  • artifacts: one row per role. The roles source, model, config and prompt are required. Other lowercase roles are allowed, once each. A path is relative to the manifest’s folder, or to --root.
  • budgets belong to the operator. null means the operator has not set the budget. Until the WER and naturalness budgets are set, every real manifest is refused (budget_unset).

What the check refuses

The check re-reads every artifact’s bytes and refuses the manifest when one of these is true. The receipt names the code and the field.

code when
consent_missing no consent record
consent_incomplete no record_id, subject, granted_on or scope; an expires_on that is not a date
consent_not_yet_granted / consent_expired the check date is before granted_on or after expires_on
consent_revoked revoked_on is set (any date)
consent_scope a profile use is outside the consent scope
unknown_use a use is outside the vocabulary
artifact_unhashed an artifact has no sha256, or the value is not 64 lowercase hex
artifact_path the path is absolute, contains .., goes through a symlink, or does not name a regular file
artifact_hash_mismatch the bytes on disk do not hash to the declared sha256
artifact_role_missing / artifact_role_duplicate a required role is missing, or a role occurs twice
seed_missing the seed is not an integer
id_missing / id_mismatch no profile_id, or it is not the id derived from the inputs
budget_missing a budget field is absent (it must be present, even as null)
budget_unset a budget is null: the detail says the operator has not set the budget
budget_owner a budget is set but set_by is not the operator (BUDGET_OWNER in the checker): the same detail
budget_invalid a budget is not a non-negative number

Also: schema_mismatch, profile_incomplete, inputs_missing and artifact_invalid for a malformed manifest.

The identity

profile_id is vp1- plus the first 24 hex characters of the sha256 of a canonical JSON object: the schema, the persona, the seed and the sorted [role, sha256] pairs (sorted keys, no spaces, ASCII). So:

  • the same inputs always give the same id, whatever the order of the artifacts, their paths or the label;
  • any change to the source, model, config, prompt, seed or persona gives a new id;
  • consent is not part of the id. Revocation makes the candidate unavailable but keeps its id. Old receipts keep the id they recorded.

derive-id prints the id for a new manifest. The id is a machine reference: a voice never reads it.

The receipt

{"schema": "voice-profile-check.v1", "manifest_schema": "voice-profile-manifest.v1", "verdict": "refuse",
 "profile_id": "vp1-...", "as_of": "2026-10-25", "artifacts_checked": 4,
 "refusals": [{"code": "budget_unset", "field": "budgets.wer_regression_max", "detail": "budget not set by <operator>"}]}

The receipt has these seven keys only. Each refusal has code, field and detail only. The receipt never holds an artifact path, a hash, the consent subject or any content bytes (audio, transcript or vector). check exits 0 for admit, 1 for refuse and 2 for a manifest that cannot be read.

Was this page helpful?