App: Live Transcription and Translation in Nextcloud Talk (live_transcription)

This app provides live transcription and translation of speech in Nextcloud Talk calls using open source AI models provided by Vosk.

The transcription is done on your own server, preserving your privacy and data sovereignty, while the translation is done using a translation task processing provider like the :ref:`translate2 app <ai-app-translate2>`. OpenAI and LocalAI integration and DeepL integration apps will soon also be supported for translation.

A good set of language models for transcription are auto-downloaded. They include Arabic, Arabic (Tunisian), Breton, Catalan, Czech, German, English, Esperanto, Spanish, Persian (Farsi), French, Hindi, Italian, Japanese, Kazakh, Korean, Dutch, Polish, Portuguese (Brazilian), Russian, Telegu, Tajik, Turkish, Ukrainian, Uzbek, Vietnamese and Chinese.

The translation capabilities depend on the installed translation task processing provider app. A list of translation-capable apps can be found :ref:`here <mt-consumer-apps>` in the "Backend apps" section.

Installation

Make sure the Nextcloud Talk app is installed.
Make sure the High-Performance Backend (latest or released after September 2025) is installed and configured in Nextcloud Talk settings. See the Nextcloud Talk install manual for more information.
Setup a :ref:`Deploy Daemon <ai-app_api>` in AppAPI Admin settings.
Install the live_transcription app via the "Apps" page in Nextcloud, or by executing

occ app_api:app:register live_transcription \
  --env LT_HPB_URL=wss://cloud.example.com/standalone-signaling/spreed \
  --env LT_INTERNAL_SECRET=1234 \
  --wait-finish

Install a Text-to-text task processing provider app for translation capabilities from the "Backend apps" section :ref:`here <mt-consumer-apps>`.

Requirements

Minimal Nextcloud version: 33
Nextcloud AIO is supported
We currently support NVIDIA GPUs and x86_64 CPUs. Only CPU-based transcription is also supported and works well on modern x86 CPUs.
CUDA >= v12.4.1 on your host system for GPU-based transcription
GPU Sizing
- A NVIDIA GPU with at least 10 GB VRAM
- 16 GB of system RAM should be enough for one or two concurrent calls
CPU Sizing
- x86 CPU with 4 threads. Additional 2 threads per concurrent call.
- 16 GB of RAM should be enough for one or two concurrent calls
Space usage
- ~ 2.8 GB for the docker container
- ~ 6.0 GB for the default models

Logs

Warnings and errors are logged to the main Nextcloud server logs.

To inspect the detailed JSON logs for lower levels, the following commands can help.

To view docker stdout/err logs:

docker logs -f -n400 nc_app_live_transcription

To view the JSON logs:

The logs are rotated when file size exceeds 20 MiB. The older log files are named lt.log.1, lt.log.2, and so on.

docker exec -it nc_app_live_transcription tail -f -n400 /nc_app_live_transcription_data/logs/lt.log

To download the JSON logs:

docker cp nc_app_live_transcription:/nc_app_live_transcription_data/logs/ /tmp/nc_app_live_transcription_logs

App store

You can also find the app in our app store, where you can write a review: https://apps.nextcloud.com/apps/live_transcription

Repository

You can find the app's code repository on GitHub where you can report bugs and contribute fixes and features: https://github.com/nextcloud/live_transcription

Nextcloud customers should file bugs directly with our Customer Support.

Limitations

The generated transcripts may not be perfect and may contain errors. It can also depend on the audio quality and the speaker's accent.
The app currently supports only a limited number of languages. More languages may be added in the future.
The languages other than English may have lower accuracy mainly due to the shipped models being smaller.
The app currently does not support punctuation in the transcription.
OpenAI and LocalAI integration and DeepL integration apps are not yet supported for translation.