Repository navigation
Use cases
Voxtrama ships a workflow for each of three kinds of recording: meetings, research interviews and lessons. Each has a workflow, a few settings worth changing and a check to make before you rely on the result. Everything on this page holds for Voxtrama 0.1 as it is shipped.
Use the Meeting decisions workflow. Its recap has key points plus the decisions, each with owner, deadline and a quote from the recording.
| Setting | Value | Why |
|---|---|---|
| Workflow | Meeting decisions | adds the Decisions section to the recap |
| Context | names of the people, acronyms, project names | Whisper spells them right, and the summary model can match an owner to a name; see Start a job |
| Summary detail | 3 (default), 2 for a short stand-up | about 8 points for a meeting, 5 for a short one |
| Recap language | the language the minutes are to be written in | it is also the language the audio is transcribed in, unless the first 90 seconds are clearly in another language |
| At the same time | when each person was recorded on their own microphone | the tracks are mixed into one before the speakers are told apart |
VOXTRAMA_MAX_SPEAKERS |
raise it to more than 4 for a larger meeting | the default ceiling is 4 speakers; see Speakers |
- In the Recap tab, press the time of each decision and listen to the quote. The owner and the deadline of a decision are the model's reading of that passage; confirm them against the audio.
- Treat every point marked Needs review as unconfirmed until you find the passage; Read and check the results shows how.
- Name the voices in the Speakers tab before you export, so that the minutes carry names and not
spk0. - Export the minutes as DOCX or PDF from Export; the file keeps each point's speaker, its time and any Needs review mark.
Use the Research interview workflow. You get key points and the themes, each with its insight and a quote.
| Setting | Value | Why |
|---|---|---|
| Workflow | Research interview | adds the Themes section |
| Context | the topic of the study, the names that recur, the technical terms | spelling of terms in the transcript |
| Summary detail | 4 for an interview with many topics | about 12 points, a short paragraph each |
| Model server | a model server on this machine | the context and the transcript reach the summary model only on the server you chose; the new-job form marks each server local or remote |
| Delete after | the period your protocol allows, 7, 30, 90 or 365 days | audio, transcript and recap are deleted then, and only an emptied record stays |
privacy: local_only |
on the generative steps, in a custom workflow | a job that would send them to a remote server is refused; see Workflows and skills |
Voxtrama does not keep a voice print: speakers are told apart inside one recording and named by hand, so no biometric profile of a participant stays in the installation. See Privacy and data.
- For each theme, press the time of its quote and listen: check that the passage says what the theme claims.
- Correct the wording of a point, or tie it to the right turn, with Edit or link to another turn; the model's own text stays in the job's record.
- Name the speakers, and give a sentence to the right person in Explore when the voices were merged.
- Export the transcript as JSON or JSONL to code it in another tool, and keep
manifest.jsonwith the study: it records the models and the settings of each job.
Use the Lesson companion workflow. The result lists key points and the key concepts, each with its definition and a quote.
| Setting | Value | Why |
|---|---|---|
| Workflow | Lesson companion | adds the Concepts section |
| Summary detail | 4 or 5 for study notes | level 5 asks for one point per minute of recording, from 15 to 40, each a paragraph with a quote |
| Recap language | the language you study in | it is also the language the audio is transcribed in, unless the first 90 seconds are clearly in another language; Same as the audio follows the lesson |
| One after another | when the lesson was recorded in several files | the files are joined in the order shown into one transcript |
| Context | the subject, the names of the authors and the technical terms | spelling of terms in the transcript |
A long lecture is read by the summary model in windows of about 12 000 characters. The model's context must be large enough for one window; Summary models and servers explains the limit.
- Press the time of a concept and listen to the explanation it comes from; confirm that the definition says what the teacher said.
- Treat a concept marked Needs review as unconfirmed.
- Use Search transcript in the Transcript tab to find the passage on a topic and play it.
- Export the recap as PDF to print, or the transcript as Markdown to keep with your notes.
This wiki is generated from docs/guide in the repository. Changes go through pull requests there.