Skip to content

v0.6.0

Choose a tag to compare

@davidamacey davidamacey released this 01 Nov 11:58
· 187 commits to master since this release

Overview

OpenTranscribe 0.6.0 is a refinement release: a refreshed interface, settings that say what
they will do, speaker data you can trust, and a processing pipeline you can compose. It is a
standard upgrade. Back up first, then update; the database migrations run on their own when
the backend starts. A few behaviours are deliberately different afterwards and are listed in
Upgrade Notes.

A refreshed interface. The gallery, navigation, file page, upload wizard, notifications and
Settings were reworked from a UX review (#739-#758) and two consistency rounds (#1243, #1244):
one modal, one form-field style, one slider, one date-range popover, one empty-state layout,
lighter type and a single accent colour. Settings opens instantly and loads each section on
demand, the app starts faster (#1131), notifications clear themselves on a retention policy
(#1217), and rate limits say how long to wait (#788).

Settings that say what they will do. Transcription and Speaker Identification are separate
tabbed sections, Privacy and Redaction are one, data tools are tabs of Data & Storage, and a new
Processing pipeline page shows where each step of the next recording runs and which
providers receive the audio (#1227).

Speakers you can trust. A speaker's name now comes only from a person or a voiceprint match,
never from an AI guess on its own. Unassigning a speaker is a correction that every automatic
path respects and that survives re-processing the file (#1233). Removing a speaker removes
every copy of its voice (#1214), and you can name a speaker inline from the transcript (#1086).

A composable pipeline. Transcription and speaker detection are chosen independently, local
or cloud, and one resolver plans every run (#1224, #1225): local Whisper with pyannote.ai,
another provider's speaker detection with any transcription (#1226), and a record on every file
of which provider and model did each step (#1229). Lite is cloud-only for transcription and
speaker detection (#1228).

Captions done properly. One server-side segmenter with broadcast rules and per-script line
breaking now drives the player's captions, subtitle downloads and burned-in video (#1244).

Translations. Every new string ships in all 12 languages, the English source copy was
rewritten so it translates cleanly (#1220), untranslated admin strings were translated with a
guard so it cannot recur (#1234), and native-level reviewers checked each locale.

Data integrity and reliability. Orphaned search documents and stored files can be found and
removed from the admin UI (#1212), a delete now leaves nothing behind in any store, durations
are a recording's real length again (#969), search indexing retries until a file is searchable
(#1182), GPU work is admitted by available VRAM and host memory with recovery from out-of-memory
(#1081, #1073), and a file is never stranded or failed because its worker died.

Security and tenancy. Tenant isolation for organizations with audited, time-boxed support
access (#1122, TENANCY_MODE), tenant-owned tags and collections (#1050, #1051), password
security levels following NIST SP 800-63B-4 (#1127), session timeouts for external sign-in
(#1106), stricter CSP, CORS and cross-origin headers (#1028-#1030), and third-party telemetry
off by default.

Upgrade Notes

0.6.0 upgrades in place like any release. The full walkthrough is in the docs:
Operations > Upgrading to 0.6.0.

  • Back up first (./opentranscribe.sh backup, or ./opentr.sh backup in a clone). The
    migration chain is one-way: rolling back means restoring that backup.
  • Amazon Transcribe now requires AWS_TRANSCRIBE_BUCKET (behaviour change, #1235). If you use
    the AWS provider and never set it, the old fallback bucket opentranscribe-tmp-<region> was
    being used; it is gone. Create a private S3 bucket in your own AWS account (same region as
    AWS_REGION), set AWS_TRANSCRIBE_BUCKET to its name and restart, or AWS jobs and the Test
    connection button will fail with a message naming the setting. The credentials must be able to
    read, write and delete objects in that bucket. If the bucket belongs to a different account than
    the credentials, also set AWS_TRANSCRIBE_BUCKET_OWNER to that account id. A deployment whose
    network has no route to STS (for example a VPC with no STS endpoint) must set
    AWS_TRANSCRIBE_BUCKET_OWNER too: without it the account id is looked up from STS, and jobs fail
    with a message naming the setting. Objects left in the
    old opentranscribe-tmp-<region> bucket, if you own it, can be deleted.
  • Twenty-three migrations, v397 to v449, run automatically at backend start. They add columns
    and tables. The data changes are small and targeted: on installs with organizations, v420,
    v422 and v430 move tags, collections and legacy speaker/vocabulary rows into the right
    tenant; v433 resets watch sources holding exactly the old 1-20 speaker default to "use my
    default"; v434 moves users who chose pyannote.ai with no key and local transcription to
    Automatic (what they already got); v436 (lite only) moves a "this server" speaker-detection
    choice to Automatic when the user's provider returns speakers. No transcript, speaker or
    voiceprint is changed by any of them.
  • Speaker labels that came only from an AI suggestion revert to SPEAKER_NN until a person
    accepts or types a name. The search-index maintenance task re-indexes the affected files in
    the background (startup, then every 6 hours, 100 files at a time); no manual reindex.
  • Unassigning a speaker now blocks it from rejoining that cluster and profile automatically,
    including after re-transcription. Name the speaker explicitly to undo it.
  • ENABLE_DIARIZATION is no longer read. Each user's Speaker detection setting decides
    (default Automatic). If you relied on ENABLE_DIARIZATION=false, choose Off in that setting.
  • Lite is cloud-only for transcription and speaker detection. Local Whisper and this
    server's speaker detection are no longer offered on lite; users with a text-only cloud model
    must pick pyannote.ai or a provider that returns speakers.
  • pyannote.ai runs when chosen. A user with a pyannote.ai key and local transcription used to
    get local speaker detection silently; pyannote.ai now runs and bills that key.
  • TENANCY_MODE defaults to auto. An install with no organizations is unaffected. An
    install with active organizations becomes multi-tenant: administrators need a support-access
    grant to open tenant content. Set TENANCY_MODE=single to keep the previous behaviour.
  • Installs that use organizations: two one-off commands after upgrading, in the backend
    container: python -m app.tasks.tenant_backfill_task (stamps older search and voiceprint
    documents with their organization, #1115; manual, also the Celery task
    tenant_backfill_organization_id) and
    python -m app.scripts.backfill_tenant_stamps --sync-voiceprints (copies v430's stamps
    onto profile voiceprints and lists ambiguous rows, #1110). Installs without organizations
    skip both.
  • Notifications auto-clear (read routine items after 24 hours, unread after 14 days, cap
    100; errors and actionable items are kept). Change it with Keep notifications for in the
    notifications panel; the choice is per browser.
  • API changes: GET /api/files/{uuid} answers 404 instead of 403 for a file you cannot
    access; DELETE /api/asr-settings/config/{uuid} and DELETE /api/asr-settings/all answer 200
    with the resulting pipeline instead of 204; GET /api/files/{uuid}/suggestions answers 204 when
    there are none yet; GET /api/system/stats no longer has diarization_enabled; the admin
    quarantine list returns owner and organization UUIDs and emails instead of integer ids. No
    endpoint was removed.
  • An AI-suggested speaker name is never the speaker's name until you accept it. A speaker's
    name in search, chat, summaries and exports now comes only from a name you typed or verified,
    or from a voice match to a profile or promoted cluster at 75% similarity or more. The AI
    speaker-identification result used to become the name on its own once its confidence reached
    75%, so a guess could be indexed and cited as a person nobody had checked. It is still stored
    and offered in the speaker editor and suggestion lists with its confidence, and accepting it
    sets the name. The speaker API's segment speaker now carries suggestion_source so the UI can
    tell the two apart. Existing libraries correct themselves: the next search-index
    maintenance passes (startup, then every 6 hours, 100 files each) re-index every completed
    file whose speaker label came from an AI suggestion; no manual reindex is needed. Legacy
    suggestions with no recorded source are treated as AI suggestions, except where the database
    proves a voice match (the speaker is linked to a profile of that name, or a stored voice match
    links it to a speaker of that name at the same similarity): the upgrade records those as voice
    matches, so they keep their name.
  • Privacy controls: nothing is deleted or rewritten on upgrade, but results can change (#1280).
    Migration v445 adds an empty organization_policy table, both deployment switches default on,
    and no stored voiceprint or gender value changes. The person-level gender switch is the
    existing Voice Attributes setting (speaker_attribute_gender_detection_enabled): anyone who
    had already turned inferred gender off now has stored inferred genders hidden everywhere,
    including from collaborators on their shared recordings (confirmed genders stay; turning the
    setting back on restores them). Four more things change: the gender migration and combined
    speaker analysis respect each owner's gender setting, the consistency repair and v4 migration
    skip scopes with voiceprints off, and a remote LLM no longer receives the speaker-name gender
    comparison. Review Settings > Privacy & Redaction > Redaction Policy after upgrading. Installs that
    use organizations: run the tenant backfill above (it is manual) before deleting stored
    voiceprints; it also stamps voiceprint documents that earlier releases rewrote without their
    organization (cluster documents, and speaker documents rewritten by the consistency repair,
    the v4 migration, a speaker merge or the index rebuild).
  • Speaker names now record whether a person or a voiceprint wrote them (#1280). Migration v449
    adds speaker.identity_source and marks as voiceprint-made only the existing identities there is
    positive evidence a voiceprint wrote: a verified speaker linked to a profile, whose name equals
    both its profile-match suggestion and the profile's name, where the profile already existed when
    the speaker was created and another speaker had been linked to it earlier (a person who types a
    name creates the profile afterwards, so a typed name that a later re-match rewrote is not
    marked). Everything else is treated as a person's, so it stays visible while voiceprints are off
    and a purge leaves it alone. A person who assigned an existing profile to a later speaker can
    still be marked, and a purge clears a marked name for good (turning the switch back on only
    un-hides). The count marked and the count of ambiguous rows left unmarked are logged at upgrade
    (v449: recognised N existing speaker identities ... left M ambiguous ones unmarked) and kept
    in system_settings under privacy.identity_source_backfill; the query that reads it and the
    comparison with unproven rows are in Operations > Upgrading to 0.6.0. Summaries and topics a
    language model already wrote are not rewritten.
  • Every speaker name, link and suggestion now records who made it, when and from what.
    Migration v449 also adds speaker.identity_source_at, identity_source_user_id (the person;
    empty for a task, and set to empty if the account is deleted), identity_source_detail and
    suggestion_detail (JSON, ids and identifiers only). A typed name, an accepted suggestion (voice
    or AI, with the provider and model), a profile assignment, a batch assignment and a cluster
    promotion or merge record the person and how; a voiceprint match, propagation or cluster join
    records the confidence, threshold, matched profile or speaker, embedding model and task; an AI
    suggestion records its provider and model and is still never applied by itself. The speaker API
    returns them read-only as identity and suggestion; the person (by) is shown only to the
    file's owner, an organization administrator and a platform administrator whose reach covers the
    file (any administrator in a single-tenant deployment; in a multi-tenant one only under a
    support-access grant, recorded like any other platform access). An account under erasure is
    shown as {"erased": true} with no id, name or email. While voiceprints are off
    they are hidden with the machine identity they describe, a purge clears and counts them
    (identity_provenance_cleared, suggestion_details_cleared), and they never reach the search
    index or a language model. Only the current state is kept; the full change history is tracked in
    #1336 (v0.7.0) and a review of the queries these columns need in #1337. The foreign key to
    user is added last in v449, so logins and provisioning are blocked only briefly instead of
    for the whole backfill; a development or --fresh database stamped at the first shape of v449
    needs a downgrade to v448 and a new upgrade (Operations > Upgrading to 0.6.0).
  • A voice-match suggestion on a speaker a person verified is no longer shown while voiceprints
    are off, is no longer recorded, and a purge clears it (#1280).
    The matchers do not write a
    voice suggestion, its confidence or its detail onto a speaker a person verified (unless the
    match replaces that identity); while voiceprints are off the suggestion block of such a speaker
    is withheld, as for any speaker, and the name the person gave stays; the voiceprint purge clears
    the suggestion columns of such a row (never its name, verified mark, profile link or identity
    record), counts them in suggestions_cleared and suggestion_details_cleared, and the residual
    check measures them. A person's accepted-suggestion record keeps only that the suggestion was a
    voice match, not its similarity score.
  • Transcript export API: GET /api/files/{uuid}/export accepts format=md and answers 400 for
    docx and pdf, which use the new asynchronous POST /api/files/{uuid}/export/document.
    Backend images are about 26 MB larger (Noto fonts). No migration.
  • update and update-full recommend a backup before upgrading. They print the backup
    command and the newest backup in ./backups before the stack is taken down, since the
    upgrade migrates the database and migrations are one-way.
  • The bundled object store is now Silo, not MinIO. The minio service runs
    docker.io/pgsty/silo (a maintained fork of MinIO), pinned by digest, on the same data volume.
    The service name, MINIO_* settings, ports and bucket names are unchanged and .env needs no
    edit, but the image changes: offline and air-gapped installs must save and load the new
    reference. Snapshot the media volume and back up the database first. Downgrading to MinIO on a
    volume Silo has used is not guaranteed safe, so rollback means restoring the snapshot.
    See Operations > Object storage and Upgrading to 0.6.0 > Object store: MinIO to Silo.

📄 This release has a large changelog. The full entry — every Added / Fixed /
Changed item — is in CHANGELOG.md.