hermine.ai Capture

Capture records your meetings, transcribes them with AI-powered speech recognition including automatic speaker detection (diarization), and generates a summary, action items, and clarification points. The workflow: record or upload an audio file, transcribe, enrich, export.

Prerequisites and activation

1. Open Settings and, under Extensions, select Apps & Plugins. Click Activate on the Capture card (see Apps).
2. Optional: install the desktop app for live recording. It is available on the Downloads page for Windows, macOS, and Linux. For file uploads only, the desktop app is not required.
3. Optional: connect Microsoft 365 for the calendar integration.

Afterwards you will find Capture in the sidebar under More apps.

Creating a meeting

1. Open Capture and click New Meeting.
2. Fill in the meeting details:
   • Title: a descriptive name (e.g. "Weekly team standup").
   • Description: optional notes.
   • Meeting Date and Location (physical or virtual).
   • Language: choose the main language or leave Auto-detect selected.
   • Meeting Type: determines how the AI analyzes the transcript (see the Meeting Types & Prompt Sets section).
3. Click Create Meeting.

Recording or uploading audio

Recording with the desktop app

1. Open the desktop app, sign in, and select the meeting.
2. Click Record. A consent checklist appears (all participants informed, consent given, company policy followed).
3. Confirm that consent has been obtained to start the recording. You can interrupt with Pause at any time.
4. Click Stop recording when the meeting ends.

Audio chunks are uploaded to secure storage continuously during the recording, so even long meetings are processed efficiently and without interruption.

Without the desktop app: uploading an audio file

You can also upload a finished recording directly in the browser:

1. Choose Upload audio file when creating the meeting or on the meeting page.
2. Select your file (WebM, MP3, WAV, OGG, FLAC, M4A, or AAC).
3. Click Upload & Transcribe.

Transcription and result tabs

Once the recording ends or the upload finishes, processing runs automatically: the audio chunks are merged, transcribed with Mistral Voxtral including speaker detection, and then enriched with AI. You see real-time status updates on the meeting page; the process usually takes 1 to 5 minutes depending on the length.

1. Content tabs Summary, Action Items (Tasks), Clarification Points, and Transcript: the AI results of the meeting.
2. Edit speakers button: speaker assignment including AI name suggestions (Apply all suggestions).
3. Export button: exports the meeting as Markdown (.md), Plain Text (.txt), or JSON (.json) (the options appear after clicking).
4. Sharing area with the Share with selection and the Share button: share the meeting within your account.

The results in detail:

• Summary: a concise overview with key discussion points and decisions.
• Transcript: each segment with timestamp, speaker, and text.
• Action items: actionable to-dos, with responsible person and due date where mentioned. Individual items can be removed.
• Clarification points: ambiguous or unresolved items that need follow-up. Individual points can be removed as well.

With Reprocess you regenerate summary, action items, and clarification points; via Reprocess with meeting type you can apply a different prompt set.

Assigning speakers (with AI suggestions)

After transcription, speakers are initially labeled SPEAKER_00, SPEAKER_01, and so on. If names were mentioned in the meeting, the AI suggests matching names ("AI name suggestions available" hint).

1. Click Edit speakers.
2. Enter names or apply the AI suggestions individually or all at once (Apply all suggestions).
3. Click Save. The transcript now shows the real names.

Calendar integration, sharing, and export

• Link calendar event: connect the meeting to a Microsoft 365 calendar event. Subject, organizer, and attendees are shown on the meeting page.
• Sharing: select a team member under Share with and click Share. Shared users get read access to transcript, summary, and action items.
• Export: download completed meetings as Markdown (.md) (full report), Plain Text (.txt), or JSON (.json) (structured data).

Meeting Types & Prompt Sets

Different meeting formats need different analyses: a standup is summarized differently from a client call or a workshop. That is what Meeting Types & Prompt Sets in the Capture settings are for:

• Each meeting type uses its own AI prompts for summary, action items, and clarification points.
• Via Manage Prompt Sets you see the system presets (e.g. for standups, project status, client calls, workshops, formal meetings, feedback conversations). You can Copy any system preset and adapt it.
• With New prompt set you create your own. Custom prompt sets are shared with all members of your account.

When creating a meeting, choose the type in the Meeting Type field; afterwards, apply it via Reprocess with meeting type.

Settings (account admins only)

The Capture Account Settings apply to all users in the account. Only administrators can access them.

1. AI Model section: model choice for summary and analysis.
2. Meeting Types & Prompt Sets section with the Manage Prompt Sets button: custom AI prompts per meeting type.
3. Data retention section with the Transcript Retention Period: transcripts are deleted automatically, audio after a fixed 72 hours.

AI model

Under AI Model you choose which language model handles the enrichment: GPT-6 Sol (default), GPT-6 Luna, GPT-5.6 Sol/Terra/Luna, GPT-5.5, GPT-5.4, GPT-5.4 Mini, Claude Opus 5, Claude Sonnet 5, Qwen 3.8 or Qwen 3.5.

Data retention

Capture minimizes the storage of personal data in line with GDPR (Art. 5 (1) (c) and (e)):

Summaries, action items, and clarification points are preserved even after the raw transcript has been deleted.

In addition, you can delete any meeting manually at any time: Delete Meeting on the detail page. Deletion is immediate and irreversible; there is no trash bin. When a user account or workspace is deleted, all associated meetings, audio files, and transcripts are removed within 30 days.

Supported languages

The Mistral Voxtral model supports German, English, French, Spanish, Italian, Portuguese, Japanese, and Chinese, among others. Alternatively, select Auto-detect.

Recording and consent

⚠️ Important legal notice

Recording meetings requires the prior consent of all participants.

• Germany: Recording conversations without consent is a criminal offense under Section 201 of the German Criminal Code (violation of the confidentiality of the spoken word) and can be punished with up to 3 years imprisonment.
• GDPR: Processing voice recordings requires a legal basis under Art. 6 GDPR.

The responsibility for obtaining consent lies with you (the user), not with hermine.ai.

Inform all participants before starting the recording and document their consent. The desktop app supports you with a built-in consent checklist before every recording start.

Desktop app API

The desktop app communicates with hermine.ai via a REST API (engine mounted at /capture):

1. Create meeting: POST /capture/api/meetings
2. Upload audio chunk: POST /capture/api/meetings/:id/audio_chunks (or POST .../presigned_upload_url for direct S3 uploads)
3. Finish recording: PUT /capture/api/meetings/:id/finish_recording
4. Check status: GET /capture/api/meetings/:id/status

Authentication uses JWT bearer tokens (identical to the hermine API v2).

Troubleshooting

Upload fails: "S3 upload: fetch failed"

If the desktop app records and produces chunks but the upload fails with "S3 upload: fetch failed", your corporate network (firewall or proxy) is usually blocking the connection to the recording storage. The audio data never leaves your machine and our storage is not reached. This is not a bug in Capture itself.

In this case, ask your IT department for the following:

• Allow outgoing HTTPS connections (port 443) to our object storage for Capture recordings: fsn1.your-objectstorage.com, including all subdomains (e.g. meeting-transcribe.fsn1.your-objectstorage.com).
• No TLS inspection for this host: the host must not be routed through a TLS-intercepting proxy (SSL inspection). This is the most common cause of "fetch failed".

If the error occurs for several people in the same company at the same time, that also points to a network-wide block rather than a local problem.

Tips

• Choose the right meeting type: a standup prompt set produces worse results for a client call than the matching type. When in doubt, reprocess with a different type.
• Name speakers early: apply the AI name suggestions right after transcription; it makes the transcript and action items much more readable.
• Use the calendar link: a linked calendar event automatically adds context such as attendees.
• Export action items: use the Markdown export to move action items straight into your project management tool.
• Review clarification points: they surface open questions and help avoid misunderstandings.

Tip: Create the meeting before the appointment and link the calendar event. Then all you need to do at meeting start is click Record in the desktop app.