# Call transcription on self-hosted Mattermost Mattermost call transcription on your own hardware with whisper.cpp by default, and what the opt-in remote API backends cost you. Source: https://mattermore.dev/transcription Project: Mattermore, a fork of Mattermost. https://github.com/dennisklappe/mattermore --- Mattermost call transcription is the feature where self-hosting actually pays off. On the default backend the transcriber does not call out to a cloud speech service. It bundles whisper.cpp and runs the model on your own machine, so the audio of your calls never leaves your infrastructure. There are remote backends too, and this page is straight about what switching to one costs you. Mattermore, a maintained fork of `mattermost-plugin-calls`, removes the Enterprise licence gate in front of the feature. What is left is infrastructure you have to run, and a choice you should make deliberately. ## The default keeps audio on your server `calls-transcriber` bundles whisper.cpp version 1.7.5, with the tiny, base and small models built into the image. You can read that off `WHISPER_VERSION` and `WHISPER_MODELS` in its Makefile. The models ship inside the container, so on the default backend there is no download step at job time and no API endpoint to configure. The consequences are the point: - No audio is sent to a third party. Not to a speech API, not to a transcription vendor, not to Mattermost, Inc. - There is no per-minute cost. You pay in CPU time on hardware you already own, not in metered API calls. - There is no vendor account, no API key to rotate, and no data processing agreement to negotiate. If your organisation self-hosts Mattermost for data protection reasons, this is the whole argument, and it is why local whisper.cpp is the backend we recommend you leave alone. A cloud transcription API would undo the reason you self-host in the first place. Running Whisper locally does not. ## The remote backends are an opt-in, and they send your audio away Local whisper.cpp is the default, not the only option. The backend is selected with the `TRANSCRIBE_API` variable on the transcriber job, and it takes three values: - `whisper.cpp`, the default. Audio never leaves the container. - `azure`, Azure Cognitive Services Speech. This one is upstream's rather than ours: the stock calls plugin already offers an "Azure AI" choice for its `TranscribeAPI` setting, next to `TranscribeAPIAzureSpeechKey` and `TranscribeAPIAzureSpeechRegion`. - `openai/api`, any remote service implementing OpenAI's `/v1/audio/transcriptions` endpoint. This one Mattermore added. It works with OpenAI itself, Groq, LocalAI, or a whisper server you run, configured through `OPENAI_API_BASE_URL`, `OPENAI_API_KEY`, `OPENAI_API_MODEL` and `OPENAI_API_TIMEOUT_SECONDS`. Now the part that matters, stated plainly: selecting either remote backend uploads the recorded call audio to a third party. Every participant's audio goes to whichever endpoint you point it at, subject to that provider's retention and training policies, and the guarantee that made self-hosting worth it is gone the moment you switch it on. That is why it is an explicit opt-in and why nothing here switches it on for you. If you want a larger model without the exposure, point `OPENAI_API_BASE_URL` at a service you run yourself. Two smaller details. The API key is sent as a bearer token and is kept out of the logs. Live captions ignore the setting entirely and always use the local whisper.cpp models. The caveat we would rather you heard from us: the OpenAI-compatible backend has only ever been tested against a mock server standing in for the API, never against a real provider. Request handling, retries, chunking and timeouts are covered by tests written against that mock. How an actual endpoint behaves is not something we have verified. Run your own trial before you rely on it. ## How it works Transcription is a job, like [recording](/call-recording). The plugin does not do the work itself. It talks to a separate service called `calls-offloader`, which launches a container for each job. For transcription that container is `mattermost/calls-transcriber`, a public image licensed Apache-2.0. `calls-offloader` is dual licensed the same way the plugin is: AGPL-3.0 for the source, MIT for compiled versions produced by Mattermost, Inc. It needs Docker access, because launching containers is how it does its work. ## What Mattermore changes, and what it does not Upstream checks `TranscriptionsAllowed()` before starting a transcription job, and that check wants an Enterprise licence. Mattermore patches the check out. It modifies AGPL-3.0 code only and leaves everything under the Mattermost Source Available Licence untouched. See [licensing](/licensing) for the full reasoning. Lifting the gate does not conjure the service. This is the sentence to take away from this page. Mattermore makes transcription permitted; it does not make it run. You still have to deploy `calls-offloader` next to Mattermost, give it Docker access, and let it pull the transcriber image. Mattermore leaves upstream's own setting defaults alone, so after installation transcription is available rather than switched on and pointing at nothing. You enable it in [configuration](/configuration) once the job service answers. The [self-hosting guide](/selfhost) covers the full stack, and [/install](/install) covers the plugin swap on an existing server. ## Choosing a model The image gives you tiny, base and small. Larger models are more accurate and slower. That trade-off is the only tuning knob most people will touch, and the right setting depends on your hardware and on how much a wrong word costs you. No accuracy figures appear on this page because none have been measured here. Run a real call through each model on your own hardware, read the output, and pick. That takes an afternoon and beats any number someone quotes at you. ## What is not claimed here Being clear about the edges is more useful than an optimistic feature list. This page does not claim GPU acceleration, live captioning during a call, speaker labelling, or support for any particular list of languages. None of that has been verified for this build. Treat transcription as something that produces a transcript from a call, and test anything beyond that yourself before you promise it to anyone. ## Is it worth it? More than recording is, in most cases, because the alternative is worse. Recording competes with someone hitting record on a laptop. Transcription competes with sending your meetings to a cloud vendor, and for a lot of self-hosted organisations that is not an option at all. The operational cost is real and it is the same cost as recording: a second service, Docker on the host, and CPU headroom for the job containers. This is a larger commitment than [group calls](/group-calls), which need one environment variable. If you already run `calls-offloader` for recording, transcription is close to free to add. If you do not, adding it for transcription alone is still a defensible trade when the audio is sensitive. ## Frequently asked questions ### Can Mattermost transcribe calls without an Enterprise licence? Yes, with Mattermore. It removes the `TranscriptionsAllowed()` check in AGPL-3.0 code. You still need `calls-offloader` running, because the licence check is not the only requirement. ### Does Mattermost call transcription send audio to the cloud? Not on the default backend. `calls-transcriber` runs whisper.cpp inside the container on your own host and no audio is sent to an external speech API. If you set `TRANSCRIBE_API` to `azure` or `openai/api` it does send call audio to that third party, which is why those are opt-in. ### What speech to text engine does Mattermost use? whisper.cpp version 1.7.5, with the tiny, base and small models built into the `mattermost/calls-transcriber` image. ### How do I transcribe Mattermost calls locally? Install Mattermore, run `calls-offloader` alongside Mattermost with Docker access so it can launch the transcriber container, then enable transcription in the plugin settings. [/selfhost](/selfhost) has the full stack. ### Does Mattermost whisper transcription support GPUs or speaker labels? Not something this page will claim. GPU support, speaker labelling, live captioning and language coverage have not been verified here, so test them before you rely on them. ### Is call transcription available on Mattermost Cloud? No. Mattermore targets self-hosted Mattermost. Cloud installations cannot load a replacement plugin. ### What does self hosted meeting transcription cost to run? CPU time and one extra service. On the default backend there is no per-minute API charge, and the recorder and transcriber images are both public and Apache-2.0. Your real cost is the host that runs the job containers. Choose a remote backend and you take on that provider's metered billing as well.