Repository navigation
FAQ
Each answer is short, and links to the page that has the details.
Yes. Voxtrama transcribes a recording with Whisper on your processor, tells the speakers apart, and asks a summary model served by Ollama for a recap, decisions, themes or concepts. Everything can run on one machine. Your first job walks through it, and Set up the summary model installs Ollama.
Yes. Voxtrama runs in Docker on your machine and serves its web interface at 127.0.0.1. It tells speakers apart inside each recording (speaker diarization) and lets you name them in the Speakers tab. See Requirements and Speakers.
Yes, once its models are downloaded. The transcription model and the speaker model are downloaded from Hugging Face the first time they are needed. After that, a job needs no network when the summary model runs on the same machine. See Privacy and data.
No, unless you configure a remote model server. Transcription and speaker diarization always run inside Voxtrama. The transcript and the context you wrote are sent out only to a remote model server, and a step marked local_only is refused before anything is sent. See Privacy and data.
Compliance depends on how a deployment is run, and Voxtrama alone cannot give it. What Voxtrama provides is facts you can assess: data stays in one folder you control, nothing leaves the machine unless you configure a remote server, no voice print is stored, and retention deletes old jobs. Privacy and data lists them. It is not legal advice.
No. Voxtrama tells speakers apart inside one recording and keeps no voice print. A voice print is biometric data, and keeping it would turn a transcription tool into an identification system. You name speakers by hand. See Speakers and Privacy and data.
Up to 4 by default. VOXTRAMA_MAX_SPEAKERS raises or lowers the ceiling. Speakers are told apart automatically and can be wrong, so check them in the Speakers tab. See Configuration and Troubleshooting.
No. Voxtrama does not use pyannote, and its code passes no Hugging Face token. The speaker model is speechbrain/spkrec-ecapa-voxceleb. See Configuration.
Yes to the first. A job's time budget is computed from the length of the audio, and an upload is limited by its size (8192 MB by default) and by the free space. For silence, a voice activity filter is always on, with no switch to turn it off, so stretches without speech are not given to Whisper. A transcript too long for the summary model is read in parts of about 12 000 characters. See Summary models and servers and Start a job.
Speed depends on your machine. As one example, with the high profile (Whisper large-v3, 4 cores per chunk, 2 chunks in parallel), 21 minutes of audio took 7 minutes 30 seconds to transcribe and 2 minutes 38 seconds to separate the speakers. Measured in September 2026 with Voxtrama 0.1.0 on one machine, an Intel Xeon W-2150B with 64 GB of RAM. Use voxtrama doctor and Troubleshooting to tune the profile and the cores.
No, not for transcription. The Voxtrama image is CPU only. A summary model can use a graphics card when the Ollama server that runs it has one. See Requirements and Summary models and servers.
Docker on macOS gives containers no access to the graphics card, and Voxtrama transcribes on the processor in any case. How fast a job runs depends on the processor, on the transcription profile and on the cores a job uses. The Efficient and Balanced profiles are faster than Maximum accuracy. See Troubleshooting.
Voxtrama 0.1 has been tested end to end on macOS with an Intel processor. Linux and Apple Silicon have not been tested end to end yet, and Windows is not supported yet. See Requirements.
It accepts audio and video files in these containers: WAV, W64, RF64, MP3, FLAC, Ogg, MOV, MP4, M4A, Matroska, WebM, AAC, AIFF, CAF, ASF, AMR, AVI, MPEG, MPEG-TS and WavPack. The limit is 8192 MB per upload by default, set by VOXTRAMA_MAX_UPLOAD_MB, and it applies to the whole request. It is Voxtrama's own limit, set by VOXTRAMA_MAX_UPLOAD_MB, and unrelated to the upload limits of hosted transcription services. See Troubleshooting and Configuration.
A summary model can, and Voxtrama checks for it. Every point of a recap carries a quote, and Voxtrama looks for the quote in the transcript. A point whose quote is not found is marked Needs review and is not shown as fact. Each point links to the moment in the audio, so you can play it back. See Read and check the results.
A step can fail. Voxtrama reads a long transcript in parts of about 12 000 characters, and a context limit lower than a part needs makes the step fail with a message that the prompt did not fit. Raise Use at most to the model's maximum, or choose a model with a larger window. See Summary models and servers.
Yes. Tick Transcription only, no summary in the Advanced section of the new-job form. The job produces the transcript and the speakers. Without any summary model, a job that needs a recap fails at that step and keeps the transcript. See Summary models and servers.
Voxtrama uses the multilingual Whisper models small, medium and large-v3, and transcribes in the language you choose or the one it detects. The recap can be written in Italian, English, Spanish, French, German, Portuguese or Dutch. See Start a job.
No. Voxtrama processes finished recordings and not a live microphone. See Your first job.
No. In 0.1, Voxtrama talks to a model server through Ollama's own API. It does not call the OpenAI API. See Summary models and servers.
No. Voxtrama works on files you already have. See Start a job.
Every job writes a manifest.json that records which models ran, with which settings, on which file. A job can be regenerated with other choices, and the steps whose inputs did not change are reused. voxtrama compare reports the differences between two manifests. See Correct and regenerate and Command line.
Delete a job from the Jobs page, or set a retention limit. Delete removes what you tick, among the audio, the transcript and the recap. Retention deletes finished jobs after 7, 30, 90 or 365 days, and voxtrama retention checks that nothing of them is left. See Privacy and data and Jobs, search and clean-up.
Copy the data folder, without models/, and start Voxtrama on the other machine. See Taking your data with you.
This wiki is generated from docs/guide in the repository. Changes go through pull requests there.