Verum Signal covers live political debates and a small number of major press conferences. To publish a claim under the name of the person who made it — a candidate in a debate, or an official holding a press conference (its principal) — the Company must determine who was speaking. It does that acoustically, using voice prints. In this policy, “covered speaker” means a debate candidate or a press-conference principal.
This policy explains what the Company creates, why, who can access it, how long it is kept, and how it is destroyed. It applies to every voice print the Company holds, wherever the covered speaker is located.
A voice print is a numeric representation of vocal characteristics — a fixed-length array of numbers produced by a speaker-embedding model. It is not a recording. It cannot be played back as audio, and it does not contain speech, words, or any content of what a person said.
The Company creates one voice print per covered speaker on the roster of an event it is covering.
During an event and in the check that follows it, the Company's systems also compute voice embeddings from the recorded audio of the event, for every voice in it (at a press conference, reporters' voices included), in order to match speech to the enrolled prints. These embeddings are biometric data in the same sense as prints. They are used only for that event's attribution and its check, and are destroyed on the schedule in Section 7.
Each print is built from at least two audio clips drawn from two independent, already-public sources — for example a campaign video and a broadcast interview, or an official's earlier briefing and an interview. The Company never builds a print from the event it is covering.
Each clip is checked by re-transcribing it against its source and confirming the expected speech appears in that clip and no other, so that a clip capturing the wrong person is caught before it becomes a print.
Source clips are audio, not voice prints. The Company retains them with the print as the evidence that the print was built from the right voice, and destroys them with it (Section 7).
A print is used only to determine which covered speaker is speaking during an event in which that person takes part. Matching is restricted to the prints of the covered speakers on that event's roster.
Prints are not used to identify anyone outside the event they were enrolled for, not used for any purpose other than speaker attribution, and not used to build any profile of an individual. Verum Signal does not produce a rating, score, or aggregate assessment of any individual person.
Moderators, panelists and audience members are not enrolled, and at a press conference neither are reporters or any speaker other than its principals. Speech is credited to a candidate when it matches that candidate's print. While a debate is under way, a line on which the match has not yet settled may be credited provisionally to the candidate last confirmed as speaking; every attribution is re-examined after the event and corrected where it was wrong. At a press conference, speech is credited to a principal only when it is in a voice the Company's system has confirmed as that principal's; other speech carries no name.
Voice prints are held by Verum Signal and are not shared. They are not sent to any third party, including the services that transcribe debate audio.
The Company does not sell, lease, trade, or otherwise profit from voice prints or any other biometric information, and does not disclose them except where compelled by law.
The Company records the public broadcast or webcast of the debates and press conferences it covers, and retains the recording and the transcript. Audio is transcribed by third-party speech services (Deepgram during the event; Rev AI for the independent check afterwards).
The Company keeps the recording because it is the evidence behind every claim published from the event. When someone disputes who said something, the recording is what allows the attribution to be checked and corrected. An event whose recording no longer exists is one whose attribution cannot be re-checked; the Company's practice is to withdraw such an event from publication rather than leave it standing.
| What | How long |
|---|---|
| Voice print | Destroyed 30 days after the covered speaker's last scheduled covered event, once the independent post-event check for that event has run and its result has been recorded. A covered speaker enrolled for several scheduled events keeps one print across them; a covered speaker who appears in a later event not yet scheduled at enrolment is enrolled again. In no case later than 45 days after the Company determines the print is no longer needed, and in no case later than 24 months after the covered speaker's last covered event. |
| Voice print, on request | Destroyed within 30 days of the request, or sooner under the row above (Section 9). |
| Source enrolment clips | Retained with the print, as the evidence that the print was built from the right person; destroyed when the print is destroyed. |
| Voice embeddings computed from event audio | Destroyed 30 days after the post-event check for that event has run. |
| Event recording and transcript (debate or press conference) | While claims from the event are published, plus twelve months. |
| Attribution test fixtures | Two recordings from past events, with the embeddings and prints derived from them, are retained as named test fixtures for validating the attribution system before changes to it. They are reviewed every twelve months and destroyed when no longer needed for that purpose. |
Every period in this table is subject to Section 12 (Legal holds).
Destruction is permanent. A destroyed print is removed from storage, including every superseded copy retained from prior enrolments, and is not recoverable. For each destroyed file, the date and time, the reason, the policy reference, the operator, and the file's size and SHA-256 digest are recorded in a destruction log that is kept separately from the prints and survives their deletion. A file is deleted, and its absence confirmed, before its log entry is committed.
A covered speaker may email privacy@verumsignal.com to ask what the Company holds about them, to receive a copy of it, to correct it, or to have their voice print destroyed.
A print will be destroyed within 30 days of the request, unless a legal hold under Section 12 applies, in which case the Company will tell the person and destroy the print when the hold is released. The person's speech at events covered after that will not be attributed to them, and depending on the event may not be covered at all: the Company publishes attribution for every covered speaker on a roster or for none of them.
To dispute a claim published under a covered speaker's name, see the corrections and disputes policy.
Voice prints, embeddings and source clips are stored only on systems the Company controls, encrypted at rest, and are not exposed through any public interface, API or repository. Access is limited to named individuals who operate event coverage. Prints are never included in code repositories, external backups, or any transfer to a third party.
If the Company determines that voice prints, embeddings or source clips have been accessed, disclosed or acquired without authorisation, it will: contain the incident and preserve the evidence; assess which covered speakers' prints were affected; destroy affected prints where continued retention serves no purpose; notify each affected covered speaker without unreasonable delay at the contact the Company holds for their campaign or office, with a description of what was affected and what the Company has done; notify any regulator whose notification rules apply; and record the incident, the assessment and the notifications in the destruction log. A print that has been compromised is not re-used.
Where an event, a covered speaker, or a claim published from an event is the subject of a dispute, demand, claim, subpoena or other legal process, or the Company reasonably anticipates one, destruction under Section 7 of the prints, source clips, embeddings, recordings and transcripts related to that matter is suspended until the matter is resolved and the hold is released. A hold applies only to the items related to that matter. The hold and its release are recorded in the destruction log, and the items are then destroyed under Section 7. Items under hold remain subject to Sections 4, 5 and 10: they are not used for any new purpose and are not shared.
Material changes will be published here with a new version number and effective date. Earlier versions are available on request at privacy@verumsignal.com.