OpenTranscribe v0.6.1: Privacy Controls, Tamper-Resistant Audit Records, and a Hardening Pass
v0.6.1 is a patch release that follows v0.6.0 quickly. Most of it is hardening and privacy plumbing: giving people and operators more control over erasure and voiceprints, making audit records harder to alter, and tightening a number of paths that handle credentials, storage and access. There is also one behaviour change for LDAP installs that you need to read before upgrading.
What is new
Erasure you can verify. Erasing an account now measures every store it touches before and after, and keeps a numbers-only receipt. An erasure counts as complete only when that check passes, and a store that cannot be measured is reported as unverified rather than as empty. After a database restore, reconciliation re-applies earlier erasures. See GDPR erasure.
Erase my account (off by default). When an administrator turns it on, users can ask to erase their own account from Settings, with a grace period (14 days by default) in which they can change their mind. See Erase my account.
Voiceprint choices per recording. The upload form can record the uploader's statement about speaker voiceprints for that file, and the owner can opt a recording out later: that stops new voiceprints from it and removes the voiceprints derived from that one recording, without deleting any speaker profile. Voiceprints can also be purged for chosen files, and an optional retention sweep (off by default) removes voiceprints that have gone unused. See Privacy controls.
Legal documents. An optional setting can require the file owner to have accepted chosen legal documents before their files are processed. Acceptance records are now kept, pseudonymised, after an account is deleted and removed after a retention period you control (2190 days by default). See Legal documents.
Token usage for every AI task. Summaries, topic extraction, speaker identification and AI redaction now each record their token usage per run, so usage reports match what was spent.
Hardening. Audit records are protected in the database (below). Protected media imports no longer pass site credentials through the task queue. Tag changes need editor access. WebSocket connections now respect sign-out everywhere and account approval. The sign-in path fails closed when an account lookup or the approval setting cannot be read, and optional-auth routes treat pending, rejected and expired accounts as anonymous. A broken personal AI configuration reports AI as unavailable instead of quietly using another provider. Storage keys are unique per file, deletes on versioned buckets remove every version, and quarantine reports honestly when storage could not revoke already-issued links.
What operators must know
- Migrations run through v457. They run automatically at startup and only go forward, so back up first. v456 builds unique indexes on object keys without blocking startup; if shared legacy keys exist, they are reported by
GET /api/admin/data-integrity/storage/keysand can be split withpython -m app.scripts.split_shared_storage_paths(see Object storage). - Audit record integrity. A database trigger now refuses deletes and rewrites of the authentication-configuration log, support-access records, erasure ledger, takedown records and legal acceptance records. On a single-role install it is an accident guard. Setting
POSTGRES_APP_USERandPOSTGRES_APP_PASSWORDruns the backend and workers as an unprivileged database role, which makes it enforceable. The new role is optional. Settings > Audit Logs shows the measured state. See Audit record integrity. - LDAP behaviour change. A plain
ldap://server address now requires StartTLS or anldaps://address; binds that would have been unencrypted are refused, and the admin connection test says so. If your directory is reached over plainldap://, enable StartTLS or switch toldaps://before upgrading, or LDAP sign-in will stop working. See LDAP. - S3 installs need three more IAM actions for version-aware deletes:
s3:GetBucketVersioning,s3:ListBucketVersionsands3:DeleteObjectVersion. See Object storage. - The first-run wizard is unchanged. Its code did not change in this release.
- The frontend and documentation images move to
nginx:1.31.8-alpine3.24. The model downloader is fixed (a cold-cache CPU run failed in v0.6.0) and now fetches every PyAnnote model the backend can load. The published v0.6.0 images contained a few stray build-machine files; they held no data and are excluded now.
Upgrade notes
./opentranscribe.sh backup
./opentranscribe.sh update
Beyond the LDAP and IAM items above, there is nothing to configure. The self-service erasure, per-file voiceprint, retention and required-legal-document features are all off until an administrator enables them. Upgrading has the general procedure.
Not measured in this release
We would rather say what we did not check. The mutation-testing ratchet for the dependency and lockout modules was waived for this release, so those modules were not mutation tested this time. The image scan has accepted findings: libxml2, plus the unchanged Debian items that upstream has marked will-not-fix. Neither is new in this release.
A security advisory covering the hardening in this release will be published separately.
Release notes and changelog
The v0.6.1 release page has the release notes, and CHANGELOG.md lists every change.
