Speech Detection

Sighthound Redactor can transcribe the audio in a video so that specific words and entire speaker turns can be found and redacted. Once a video has been transcribed you can search the transcript, seek to any word in the video player, and redact words one at a time or as a selection. Redactor can also detect who spoke when (speaker diarization), so you can redact everything a single person said in one action.

Supported Languages

English ships with Redactor and is ready to use straight away. When your organization enables it, the following additional languages can be downloaded on demand the first time you select them (see Downloading a Language):

Czech
Dutch
French
German
Italian
Polish
Portuguese
Russian
Spanish

Chinese, Japanese, and Korean can be transcribed but use timing-only alignment — there is no word-level alignment model for these writing systems — so they are not offered in the editor's language picker.

Opening the Speech Panel

The Speech button is in the bottom right of the video editor. If the current file has no audio track, the button is disabled.

Each click cycles the panel:

Side panel — the transcript opens on the right of the video.
Overlay — the transcript covers the video area (the playback controls stay visible). Overlay transparency can be adjusted in Speech Display Options.
Hidden — the panel closes.

The panel is only open for the current video. Closing the video or opening another one starts with the panel hidden again.

Transcribing Videos

Open the speech panel and choose a language from the picker. If that language has not been transcribed yet, Redactor first shows a Detect dialog:

Speaker Detection — leave this on to identify who spoke when. Turn it off to transcribe words only, with no speaker colors, names, or speaker search. Your last choice is remembered the next time you transcribe.
Automatic — Redactor estimates how many speakers are in the audio (only when Speaker Detection is on).
Speaker count hint — enter how many speakers you expect, from 1 to 16, when you already know the number.

Click Start to begin. A "transcribing speech" screen appears with a loading bar. You can wait until it finishes, or click Run in background if there are other projects you'd like to look at.

When transcription is done you can search, seek to, and redact words — and, if Speaker Detection was on, work with the speakers Redactor detected.

A check mark next to a language in the picker means that language has already been transcribed for this video. Choosing it switches the transcript without running detection again.

If you pick a language that hasn't been downloaded yet, Redactor will offer to download it first.

Downloading a Language

The first time you choose a language that isn't already installed, Redactor shows a confirmation dialog with the approximate download size. Confirm, and the language model is fetched in the background while a progress indicator keeps you informed. You can cancel the download at any time; if you do, nothing is left behind and you can try again later.

Downloaded languages persist — they stay available for future videos and survive application updates and reinstalls, so each language only needs to be downloaded once. Only one language downloads at a time. After the download finishes, choose the language again to transcribe it.

Reading the Transcript

The transcript is laid out like an editor:

Each line starts with a timestamp. Click the timestamp to seek the video to that line.
Click a word to seek the video to that word.
During playback, the word being spoken is highlighted. If the transcript is long enough to scroll, an arrow button in the toolbar turns Follow playback on or off so the view can stay with the current word.
When Speaker Detection is on, each speaker can be shown with a name pill and colored (or grayscale) text — see Speech Display Options.

The toolbar above the transcript also has:

The search field and a ? button that opens a short search-syntax reference.
The language picker. A cloud icon means that language still needs to be downloaded.
A download button to export the transcription (after a language has been transcribed).
A cog button for Speech Display Options.

Speech Display Options

Open the cog in the speech toolbar to change how the transcript looks. Changes preview immediately; Save keeps them, Cancel restores the previous settings.

Show speaker name at the start of each line — shows a name pill (for example, Speaker #1 or a name you assigned) at the start of each speaker's turn.
Speaker text coloration — None, Grayscale, or Color. Color and grayscale tint each speaker's words so you can see who is talking.
Line break pause (seconds) — start a new line after a pause at least this long.
Start a new line on every speaker change — when on, a new speaker always begins on a new line. When off, speakers can share a line and a name pill still appears at each change.
Maximum gap between words to merge (seconds) — when you redact several nearby words, pauses shorter than this are included in one muted region instead of leaving short audible gaps.
Overlay transparency (%) — how see-through the transcript is when it is shown as an overlay on the video.

Searching the Transcript

The search field at the top of the panel is a live filter. As you type, the transcript shrinks to the words that match. Click a remaining word to seek the video to it. Clear the field to show the full transcript again.

Search looks at each word on its own. If you have corrected a word, search uses the corrected text.

Click ? next to the search field to open the same syntax reference in the app.

Plain text

Type any text for a case-insensitive partial match:

hello — words that contain “hello”
hello world — words that contain both “hello” and “world”

Because each search term is matched against a single word, a two-word phrase will not find those two words sitting next to each other unless one word happens to contain both strings.

Regular expressions

Wrap a pattern in re(/ and /):

re(/^h/) — words that start with “h” (not case sensitive)
re(/a\/b/) — words that contain the literal text a/b (write \/ for a slash inside the pattern)

Speakers

When Speaker Detection found speakers, you can filter by who spoke:

s(1) — words from Speaker #1
s(Officer) — words from the speaker named “Officer” (not case sensitive)

Names with spaces work inside the parentheses: s(Bob Builder).

As soon as you type s(, a list of speakers appears so you can pick one instead of typing the full name. Arrow keys move through the list, Enter or Tab inserts the highlighted speaker, and Esc closes the list.

If Speaker Detection was turned off, or the current language has no speakers, s(...) shows an error.

Excluding matches

Put ! in front of any of the above:

!hello — words that do not contain “hello”
!re(/foo/) — words that do not match the pattern
!s(1) — words not from Speaker #1

Combining searches

You can put several pieces in the same search, separated by spaces. In most cases a word must satisfy every piece:

re(/^h/) s(Officer) !yo — starts with “h”, spoken by “Officer”, and does not contain “yo”

The exception is several speaker filters without !: s(1) s(2) means Speaker #1 or Speaker #2. A word can only belong to one speaker, so requiring both would never match anything.

When search shows an error

A red message appears under the field and the transcript shows no matches until the search is valid. Typical causes:

An unfinished pattern, such as re(/foo without the closing /)
An unknown speaker number or name
A lone ! with nothing after it

A valid search that simply matches nothing shows No matches found.

Redacting Words

You can redact words by scrolling through the transcript or by searching first so only the words you care about remain. Right-click a word (or a selection) and choose:

Redact — mute that word, or every word in the selection.
Unredact — remove those redactions.

If the selection mixes redacted and un-redacted words, both items are offered. Selected redacted words use a distinct highlight, and adjacent selected redacted words read as one continuous band.

Redacting mutes the matching span of audio; the redacted word is highlighted in the transcript and a muted region appears on the audio timeline. Nearby words can be merged into one muted region using the merge-gap setting in Speech Display Options.

All of these actions can be reversed with Ctrl+Z from inside the panel.

Speakers

When Speaker Detection is on, Redactor also performs speaker diarization — it works out who spoke when and groups the words by speaker. Speakers are numbered in the order they first appear (Speaker #1, Speaker #2, and so on) until you give them a name.

Each speaker can have a name pill and colored or grayscale text so you can see at a glance how a conversation moves between people. Coloration never hides the redaction highlight.

If Speaker Detection was off, the transcript is words only: there are no name pills, no speaker colors, and no speaker items in the right-click menu.

Redacting by Speaker

Diarization lets you redact everything a person said without picking through the transcript word by word. Right-click any word that belongs to a speaker and choose:

Redact speaker #N — mute every word spoken by that speaker. If you have named the speaker, the menu uses that name instead of the number.
Unredact speaker #N — remove those redactions again.

When you redact a speaker, consecutive words from that speaker — including the short pauses between them — are merged into a single continuous muted region, so the speaker's turn plays back as one uninterrupted silence instead of a string of choppy micro-gaps. Unredacting removes exactly the same region, leaving nothing behind.

Selecting by Speaker

With a word selected, right-click to work with whole speakers at once:

Select speaker #N — select every word spoken by that speaker (again, a name is used when you have assigned one).
Select all other speakers — select every word not spoken by that speaker.
Invert selection — swap what is selected with what is not.

You can then redact, correct, delete, or reassign the selection in one step.

Reassigning & Creating Speakers

Automatic diarization is highly accurate, but occasionally a word or run of words is attributed to the wrong speaker — or you may simply want to give a speaker a meaningful name, such as “Officer” or “Witness”. Select the word or words in question, right-click, and choose Reassign Speaker to open the reassignment dialog.

You can also right-click a speaker name pill to reassign just that turn on the line.

Start typing in the box and the list of existing speakers is filtered as you type. From there you can either:

Choose an existing speaker — the selected words are reassigned to that speaker. The button reads Reassign Speaker.
Type a new name — a new, named speaker is created and the selected words are assigned to it. The button reads Create New Speaker.

Speaker names follow a few simple rules:

They cannot contain the characters (, ), or !.
Leading and trailing spaces are removed automatically.
They can be at most 64 characters long.
They must be unique — a name that already belongs to another speaker cannot be reused (matching ignores upper/lower case).

Reassigning or creating a speaker updates the name pills and colors throughout the transcript, and can be undone with Ctrl+Z.

Relabeling Speakers

To rename a speaker without moving any words, right-click a word (or that speaker's name pill) and choose Relabel Speaker. Enter the new name and confirm. The same name rules as reassigning apply. Relabeling can be undone with Ctrl+Z.

Correcting Words

If a word was mis-heard, right-click it and choose Correct Word. The dialog shows the original word, the speaker (when known), the timestamp, and the duration. Edit the text and click Done.

To correct several words, select them first and choose Correct Selection. Use < and > to move through the selection; Done saves every change you made in the dialog.

Search and the transcript both use the corrected text. You can still see the original word in the correction dialog.

Deleting Words

Right-click a word and choose Delete word, or select several words and choose Delete selection. Deleted words are removed from the transcript only — the audio itself is unchanged unless you also redact it. Deletion can be undone with Ctrl+Z.

Exporting the Transcription

After a language has been transcribed, click the download button in the speech toolbar and choose a format:

Raw Text — every original word joined by spaces, with no line breaks or speaker information.
Plain text — one line per speaker run or long pause, without timestamps.
Plain text with timestamps — the same layout, each line starting with a date and time.
JSON — the full speech data, including speaker labels, corrections, and every word.

On the desktop app the button saves the file; in the browser it downloads it.