How to Transcribe Audio Offline

Download the sample

Sample data · Two speakers discuss a design review and engineering work. Download the WAV (0.8 MB), read the reference text and review notes, and compare the Markdown notes.

1. Prepare the app and models while online

ZoneVoice is an invitation-based preview. Request early access and use the preview build supplied for your platform. In Settings → Local models, install the transcription and notes models, then verify their status. Install speaker models too if you want speaker separation.

Whisper small: transcription.
Gemma 4 E4B: summaries, tasks, and mind maps.
Speaker diarization: optional separation of voices within the file.

Wait until the required models show Ready before going offline. Model and runtime downloads require connectivity; the imported audio does not need to be uploaded. The desktop package manages its native dependencies, so this workflow does not require you to install Python, FFmpeg, Ollama, or Docker.

2. Import the recording and process locally

Open a workspace, select Add recording → Choose file, and choose the sample WAV. Keep Process with local models on. Expand Processing settings to check language, template, and Identify speakers, then choose Import and process.

Import dialog with the synthetic WAV and local processing enabled — Choose an audio or video file, then select local processing and speaker identification.

Once the required models show Ready, you can disconnect from the network and process the file.

3. Check the transcript and speaker groups

The run produced 12 timestamped segments and two speaker groups. The voices remained Candidate 1 and Candidate 2: a temporary group is not a confirmed person’s identity. Check names against the audio before confirming them.

Completed local transcript with timestamped segments and two candidate voices — Review the speaker labels and use timestamps to replay the original words.

4. Check the summary against the surrounding words

The notes included two key points, two tasks, two chapters, and a mind map. The two tasks retained Friday and Monday. Both owners are marked Not specified. Check the surrounding transcript before assigning either task.

Generated summary with Friday and Monday tasks, both with owner Not specified — The proposal request does not name an assignee. The second speaker introduces himself as Daniel and makes the technical-notes commitment, but the generated owner field did not capture that.

Clicking the technical-notes task’s citation opened the transcript and moved the player to 00:21. Its “Monday” deadline appears in the following segment at 00:24. The Friday task’s citation points only to the deadline word at 00:11; the preceding segment contains the task itself. Read or replay the surrounding passage, not just the linked word.

Task citation opens the transcript with playback positioned at 00:21 — A valid timestamp establishes a location. It does not establish that a generated claim or owner is correct.

5. Export and inspect the same result

Open the recording’s actions menu and choose Export → Complete notes · Markdown. Save the file, then inspect its summary, tasks, chapters, mind map, and timestamped transcript. The downloadable result here includes the summary, tasks, and transcript shown above.

Other export options include transcript text, SRT subtitles, mind-map Markdown, JSON backup data, and the original recording. Exporting to a cloud-synced folder lets that folder’s service upload the file.

Use the notes across recordings

Continue with the meeting knowledge base guide to track decisions across interviews, or see how MCP gives an AI client access to saved notes.

How this offline example was checked

The sample uses synthetic voices. It was processed in ZoneVoice 0.1.0 on an Apple silicon Mac with the required models already installed. External network access was blocked for the app and its model processes; local connections between those processes remained available. Transcription, speaker separation, notes, and Markdown export completed.

This check covers file processing after model setup, not a new installation or transcription accuracy on real meetings. The run details include model information and file hashes. The downloadable Markdown is the unedited app export.