Video and Audio Transcription – DekkoDEMS knowledge base

DekkoDEMS can transcribe the speech in your audio and video files into a searchable, speaker-attributed transcript. Transcription runs on Dekko EdgeAI, an AI model that executes entirely in your browser, on your own device. The audio never leaves your device, and no audio or transcript text is ever sent to a transcription service. Only the resulting transcript is attached to the file inside your case, end-to-end encrypted like everything else in DekkoDEMS.

This makes transcription suitable for sensitive evidence such as interviews, interrogations, bodycam and dashcam footage, and recorded calls, without handing your media to a third-party cloud service.

💡

The transcript is an AI-produced draft. Dekko EdgeAI may not be 100% accurate. Every transcript, along with its speaker labels, sound detections, and scene read-out, must be reviewed thoroughly against the original recording before you rely on it.

How it works in brief

When you transcribe a file, DekkoDEMS downloads the Dekko EdgeAI model to your browser, then runs the audio through it locally. The model:

transcribes speech with timestamps

separates the different speakers

detects non-speech sound events (sirens, dogs, vehicle horns, and similar), each with a timestamp and a confidence score

gives a short read-out of the background scene or ambience

Everything in this sequence happens on your device. The only thing that travels over the network is the one-time model download. Because each of these outputs is produced by an AI model, each of them can be wrong and needs to be checked.

The Dekko Edge AI model is downloaded once and reused

The first time you run a transcription, your browser downloads the Dekko Edge AI model from Dekko infrastructure. This is a one-time download, and you will see a progress bar ("Downloading Dekko Edge AI model").

Once downloaded, the model is stored in your browser for future use. It does not need to be downloaded again on the same browser and device. Later transcriptions reuse the stored copy and start much faster, with no further download.

The model download is the only network activity in the whole process. Your audio and the transcript are never uploaded. The model comes to your device, not the other way around.

💡

Note: Because the model is cached per browser, the first transcription on a new browser, a new device, or after clearing your browser data will download it again.

Who can transcribe

Only the file owner or a case manager with permission to use EdgeAI features can transcribe a file. The action appears on any file DekkoDEMS recognises as audio or video. The action appears in the file's ··· menu on any file DekkoDEMS recognises as audio or video.

Transcribing a file

Upload your audio or video file to a case as you normally would. When the upload completes, DekkoDEMS offers to transcribe straight away.

In the Transcribe media file(s)? dialog, choose Transcribe Now, or Skip For Now* and start it later. Your upload completes either way.

If you choose to translate the file, a translated transcript will be created. See the translation article for more information.

Transcription runs at the same time as your upload and may keep running after the upload finishes. The language is detected automatically, so there is no language to set.

💡

Keep this tab open. Transcription runs on your device, so closing or reloading the tab stops the run and the in-progress transcript is lost. You can keep working in other tabs while it runs.

You can queue more than one file, and they transcribe one after another. A Cancel Transcription button is available while a run is in progress.

* If you have already uploaded a file or want to transcribe later, find the Transcribe option under the More dropdown:

The transcript viewer

Once a transcription finishes, the file shows a Transcript badge, and you can open the transcript from the file. The viewer gives you:

A timestamped, speaker-attributed transcript. Click any line to play the recording from that point.

Editable speakers. Speakers are labelled Speaker 1, Speaker 2, and so on. Click a speaker name to rename it (for example, to a real name) or to merge speakers.

Editable lines. You can correct a wrong word or reassign a line to a different speaker.

Search. Search the transcript text to jump to a moment.

Detected sounds. A list of non-speech events, each with a timestamp and confidence, that you can verify against the recording.

A scene/ambience read-out (for example, "Background: traffic, likely outdoor/urban").

Export. Export the transcript to the case or as a plain .txt file for briefs, disclosure bundles, or court prep.

Editing the transcript only changes the transcript. It never alters the underlying recording. The viewer is designed for review: it puts the text next to playback so you can confirm every line against what was actually said.

It is an AI draft and must be reviewed

The transcript is labelled a Dekko Edge AI draft, and that label is important. Like any AI-produced output, it may contain errors and may not be 100% accurate. Before you rely on a transcript, share it, or use it in any official capacity, review it thoroughly against the original recording.

In particular:

Words can be wrong or missing. Misheard words, dropped words, and incorrect punctuation are possible, especially with crosstalk, background noise, accents, or poor audio quality.

Speaker labels are estimates. Who said what, and how many speakers there are, is the model's best guess and should be confirmed.

Sound detections are estimates. Each detected sound carries a confidence score. Treat these as prompts to go and listen, not as findings.

The scene/ambience read-out is an estimate. It is a high-level guess about the environment, not a verified fact.

The reviewer confirms; the model only suggests. Do not treat any part of the transcript as accurate until a person has checked it against the source recording.

Privacy and the audit trail

On-device. All recognition and analysis run in your browser, on your hardware. Audio and transcript text are never transmitted to a transcription service.

End-to-end encrypted. Files stay encrypted, and chunks are decrypted on-device only to process or play.

Non-destructive. Editing a transcript never changes the original recording.

Auditable, metadata only. Creating, viewing, exporting, and editing a transcript (line edits and speaker renames) are written to the file's audit log, recording only counts, language, speaker labels, and timings, and never the transcript text itself. You can open the file's audit log to see this history, including before and after details for each line edit.

One transcript per file. Re-transcribing overwrites the previous transcript and resets manual speaker labels.

Good to know

Formats. Any audio or video file (for example .mp3, .mp4, .webm). The action appears wherever DekkoDEMS recognises a file as audio or video.

Length. Short, high-quality recordings transcribe in a single pass. Longer or larger files switch to a streaming path that processes the audio a window at a time, so there is no hard length limit, but long recordings take several minutes.

Speed varies by device. A device with a GPU is markedly faster. The slower fallback runs the same model at the same accuracy, just more slowly.

Cancelling discards progress. Once a run is cancelled, the in-progress transcript cannot be recovered.

No translation. The transcript is produced in the spoken language. Dekko Edge AI transcribes, it does not translate.

FAQ

Does any audio or text get uploaded to transcribe it? No. Recognition runs entirely in your browser. The only thing downloaded is the Dekko Edge AI model, once, and that stays stored in your browser for reuse.

Is the transcript accurate? It is an AI-produced draft and may not be 100% accurate. Always review it thoroughly against the original recording before relying on it, and confirm speaker labels and sound detections rather than taking them at face value.

Why is my transcription slower than a colleague's? Speed depends on your device. A machine with a GPU uses the faster path. Without one, the same model runs at the same accuracy but slower.

Can I fix a wrong word or speaker? Yes. Lines and speaker names are editable in the viewer. Edits change only the transcript, never the original recording.

Can I get the transcript out of DekkoDEMS? Yes, export it as a .txt file from the viewer.

Why do I have to download the model again? The model is cached in your browser. A fresh download happens on a new browser or device, or after your browser data is cleared.