Skip to main content

OpenTranscribe v0.6.0: A Friendlier Interface, Speakers You Can Trust, and a Pipeline You Compose

· 12 min read
OpenTranscribe Team
OpenTranscribe Development Team

Hey everyone. We have been heads-down building for the past several weeks, and v0.6.0 is the result. This release is about how OpenTranscribe feels to use. If you read our earlier release posts, you will notice they were mostly about the engine: faster transcription, better speaker detection, more languages, more providers. This time we stepped back from the engine and asked a different question: is the app itself easy, intuitive and modern to use?

Below is what we have been building, why we did it, and what changes for you. When you are ready, try it out, and the full release notes and changelog are linked at the end.

Why this release is different​

Until now, most of our effort went into the backend: the processing pipeline, the models, the search, the integrations. That work matters, but a transcription tool people cannot find their way around is not much use, however good the transcripts are.

So for v0.6.0 we made the interface a first-class goal. We partnered with a professional UI designer who looked at the app the way a new user would and gave us direct, specific feedback: where it was confusing, where it was inconsistent, where it looked dated, and where a setting did not tell you what it would do. It was not always comfortable to read, and it made the app better. We worked through that feedback screen by screen, and then did two more rounds to catch the inconsistencies the first pass exposed.

We also wanted the people who run OpenTranscribe to be able to trust it. That is why the other half of this release is about speakers, privacy and reliability: what the app tells you should be true, and what you switch off should stay off.

A refreshed interface​

If you open v0.6.0 after using an earlier version, the first thing you will notice is that it looks calmer and more consistent.

Every dialog now shares one design, every input has one style, every slider looks the same, and there is one accent colour instead of several. The type is lighter, every menu item has an icon, and the date-range picker works the same everywhere. Small things, but together they make the app feel like one product instead of a collection of screens.

A few places got more than polish:

  • The gallery has a tidier two-row toolbar, a filter panel you can collapse and type dates into, a filter by owner, and a status indicator that never gets clipped.
  • The file page is built around the thing most people do there, which is fixing and naming speakers. Comments sit under the transcript, and a Jump to current playback position button replaces the minimap that could accidentally rewind your audio.
  • Search has a full-width query bar, details you can pin, and matches you can collapse. There is a short guide in Search and filters.
  • Navigation is down to the four places you actually go: Gallery, Transcript Search, AI Chat and Speakers. File status moved into Settings.
  • Notifications now show each file's progress in three plain stages (Main processing, Finishing, AI enrichment) and only say a file is done when it really is. Downloads get one live row, and the panel tidies itself up on a retention policy you choose. See Notifications and the About dialog.

It is also faster. The app loads much less code at startup, Settings opens instantly and fetches each section when you open it, and when you hit a rate limit the message tells you exactly how long to wait.

Settings that tell you what will happen​

One piece of the designer's feedback stuck with us: a settings page should never make you guess. So we reorganised it. Transcription and Speaker Identification are separate tabbed sections, Privacy and Redaction are together, and data tools live under Data & Storage.

The new Processing pipeline page is the one we are happiest with. Before you upload anything, it shows where each step will run, which providers will receive your audio, and anything that would stop it, with a link to the tab that fixes the problem.

Speakers you can trust​

A speaker's name appears in search, chat, summaries and exports, so it had better be right. Starting with this release, a name only comes from two places: a person who typed it, or a voiceprint match at 75% similarity or higher. AI name suggestions are still made, and you still see them with a confidence score, but they no longer quietly become the name.

After you upgrade, speakers whose names only ever came from an AI guess go back to SPEAKER_01 and friends until you accept a name. This happens in the background, with no reindex. Each speaker also remembers whether its identity came from a person or a voice match, so you can always tell which is which.

Corrections now stick. Unassigning a speaker from a cluster means "this voice is not that person", and nothing automatic puts it back, even if you re-transcribe the file. And you can finally name a speaker straight from the transcript's speaker menu.

You control the biometric-style features​

Recognising the same person across recordings, and inferring a speaker's gender from their voice, both work on sensitive data, so each now has a switch. The switches are restrict-only: the instance, an organization (managed edition) and each person can turn a feature off, and if any layer says off, it is off. If the policy cannot be read, the answer is off.

Turning a switch off hides existing data and stops new data. It deletes nothing, and speakers are still told apart within a recording. Deleting is a separate, deliberate action: an explicit purge with a typed confirmation, recorded in the audit log, that reports how much it removed from each store and then checks that nothing is left. Both switches default to on, so upgrading deletes and rewrites nothing. The one exception is anyone who had already turned inferred gender off in Voice Attributes: those stored genders are now hidden everywhere, and any gender a person confirmed is always kept.

Wherever a gender is shown, the interface now says it is inferred from voice and links to the switch. The switches are under Settings, Privacy & Redaction, and the full reference is in Privacy controls.

A pipeline you compose​

Transcription and speaker detection used to travel together. Now they are two independent choices, each local or cloud, and one planner works out every run. Local Whisper with pyannote.ai speaker detection works, and so does another provider's speaker detection on top of any transcription. Every file records which provider and model finished each step, and the new Processing section on the file page shows it.

Super admins can manage local models from Settings, Transcription, Local models: download any Whisper model on the shipped list with a disk-space check, live progress and cancel, then switch the active one. Only the Whisper workers restart, and only after running jobs finish. You can also limit which cloud providers regular users may choose. See Local transcription models.

The lite image (no GPU) is now cloud-only for transcription and speaker detection, and still computes voiceprints on your own server. Deployment variants lists every valid combination, generated from the code.

Export in the format you need​

The Export menu now covers TXT, JSON, CSV, SRT, VTT and Markdown, plus Word (.docx) and PDF. Word and PDF are built on the server with numbered lines you can cite by page and line, right-to-left layout for Arabic and Hebrew, and proper shaping for Japanese, Thai and other scripts. Your redaction policy is applied while the document is built, so it cannot be bypassed from the browser. The Markdown export adds front matter for note-taking tools. See Document export.

Captions follow broadcast conventions: two lines, a comfortable reading speed, a speaker's name at the start of each turn. The same captions drive the player, subtitle downloads and burned-in video, so what you watch is what you export.

For teams (managed edition)​

The managed edition gets tools for teams that need governance: restricted collections, legal holds on files or whole collections, an organization audit log you can filter and export, and a takedown queue. These are switched off in the open-source build. Everyone does get content-access auditing (who opened a recording or exported a transcript, off by default) and a durable record whenever a file is quarantined. See Support access.

Under the hood​

Less visible, but just as important to us:

  • Recovery: a file is never stranded or failed because its worker died, and stale "in progress" rows retire themselves.
  • GPU work is admitted by available memory and recovers from out-of-memory errors.
  • Data integrity: the admin UI can find and remove orphaned search documents and stored files, and a delete now leaves nothing behind in any store. Durations are a recording's real length again.
  • Security: tenant boundaries are checked in every layer that reads them. Sign-in follows NIST SP 800-63B-4 password levels, with an optional breached-password list, session timeouts for external sign-in, stricter browser headers, and third-party telemetry off by default. Dependencies are updated, and release images are rebuilt with refreshed OS packages and scanned before they are published.
  • Search quality, honestly reported: we are still measuring how much chat reranking helps. The numbers, including the unflattering ones, are in RAG evaluation.

And as always, every new string ships in all 12 languages. The English source copy was rewritten so it translates cleanly, and native-level reviewers checked each locale.

One change you should know about: the object store​

OpenTranscribe keeps your media in an S3-compatible object store, which has been MinIO. MinIO removed its public images, its community repository is archived, and its last community release line carries published high-severity advisories with no community fix. A fresh install could no longer even download it.

So the bundled store, the minio service, now runs Silo, a maintained fork of MinIO. It keeps MinIO's settings, routes and on-disk format, so your .env, ports, buckets and data volume stay as they are and no data is copied. We checked this against real encrypted data from an earlier install before shipping it. The reasoning, the alternatives we weighed and what other projects chose are in the decision record and Object storage.

Before you upgrade, back up the database and snapshot your media volume if you can, because going back to MinIO on a volume Silo has used is not guaranteed safe. Air-gapped installs need to load the new image. Steps, checks and rollback are in Upgrading to 0.6.0.

Try it​

The upgrade is a standard one. Back up first:

./opentranscribe.sh backup
./opentranscribe.sh update

The database migrations run automatically when the backend starts, and they only go forward, so the backup is your way back. A few behaviours are different afterwards: AI-only speaker labels, unassignment, ENABLE_DIARIZATION, lite, tenancy mode, notification retention, and Amazon Transcribe now needs AWS_TRANSCRIBE_BUCKET. Upgrading to 0.6.0 walks through each one and what, if anything, you need to do. New here? Start with the Docker Compose install guide.

We would love to hear what you think, especially about the new interface. Open an issue or a discussion on GitHub, and tell us what still feels clumsy. This release came from honest feedback, and we want more of it.

A note on where we are​

OpenTranscribe is still a side business and a side project that we are working to grow. It is built in the hours around other work, which is why releases take as long as they do and why we would rather ship something solid than ship on a schedule. Nothing about that changes how seriously we take your data or your upgrade, but it does mean we rely on you more than a large team would.

If OpenTranscribe is useful to you, the most helpful things you can do are to try it and tell us honestly what works and what does not, report bugs with enough detail for us to reproduce them, star the repository, tell a colleague, and, if you are comfortable with the code, read the contribution guidelines and send a fix. Every one of those helps this project grow, and a good share of what is in v0.6.0 started as a user's report. Thank you for being part of it.

Release notes and changelog​

The v0.6.0 release page has the release notes, and CHANGELOG.md lists every change with issue numbers.