What we do — Harvan Data Labs

Capabilities

What the pipeline handles

Send audio only, video only, or media with existing transcripts. Each layer can be run on its own or together.

Audio redaction

Anything personal that is said aloud comes out of the audio. The recording keeps its original length and timing, so it still lines up with transcripts and annotations. Where the voice itself identifies the speaker, we can mask that too while keeping the speech intelligible.

Names and nicknames
Phone numbers
Account and card numbers
Addresses
Dates of birth
Employer names on request

Result: Every spoken personal detail removed from the audio
Preserved: Duration, timing, speaker turns, room tone
Optional: Voice masking that keeps speech intelligible
Channels: Mono and dual-channel (agent/customer) handled separately

Video redaction

Faces are obscured for the whole time a person is on screen — they stay unidentifiable as they move, turn or leave and re-enter the shot. Screens, documents, name badges and burned-in captions can be covered in the same job.

Faces, tracked frame by frame
Name badges and ID cards
Screens and documents in shot
Burned-in text and captions
Backgrounds on request

Result: Nobody in frame is identifiable, in any frame
Preserved: Frame rate, resolution, audio sync
Optional: Body or background masking, cropping, letterboxing
Handles: Movement, partial profiles, more than one person in frame

Transcript redaction

The same details are replaced in the text with neutral tags, at the same timestamps as the audio and video, so all three stay consistent. Send your existing transcripts and we will redact those instead of producing new ones.

[NAME]
[PHONE]
[ADDRESS]
[ACCOUNT]
[EMAIL]
[DATE_OF_BIRTH]

Result: Text carries no personal details, still aligned to the media
Input: Our transcripts, or yours in JSON, SRT, VTT or plain text
Consistency: The same speaker keeps the same pseudonym across a file
Output: Your schema, or ours

Transcription and preparation

If the redacted media is headed for training, we can deliver it ready to ingest: verbatim transcripts with speaker labels and timestamps, segmentation, speaker separation and QA of the media against its metadata.

Verbatim transcripts
Speaker labels
Utterance timestamps
Segmentation
Speaker separation
Metadata QA

Languages: English, Indian English and Indian languages
Output: JSON, SRT, VTT or plain text
Structure: Folder layout and manifests to match your loader
QA: Sampled human check against the audio

Layer	What is detected	What happens to it
Video	Faces, on-screen text, name badges, screens and documents in frame	Obscured for every frame they appear in; cropped out instead, on request
Audio	Names, phone numbers, addresses, account and card numbers, dates of birth, organisation names on request	Removed from the audio, with the original timing preserved
Voice	Speaker identity	Optionally masked, so the speaker cannot be recognised but the speech stays usable
Transcript	The same details, at the same points in the media	Replaced with neutral tags such as `[NAME]` `[PHONE]` `[ADDRESS]`
Metadata	File names, device IDs, GPS and embedded tags	Stripped and re-issued under neutral IDs
Review	Every redacted file	Checked by a person, with a log of what was removed and where

What the pipeline handles

Audio redaction

Video redaction

Transcript redaction

Transcription and preparation

What gets removed, layer by layer

Not sure it covers your material?