False start detection

Every "wait, let me say
that again"
— caught.

Sapari's AI reads your transcript and identifies when you restarted a sentence, trailed off, or redid a line. Each one becomes a reviewable card on the timeline.

Start free trial

7 days · 30 AI minutes · No credit card

Transcript · 01:12
False start · 0.92

So the way I think about this is actually, the way I think about this is you need to decide what matters first.

Detected
Restart fragment
Final take

The problem

Silence removal
doesn't catch these.

The audio is full and the transcript has words. But those words aren't the take you're keeping. Finding them manually means scrubbing and reading the transcript — exactly the post-production work you're trying to avoid.

45-min recording → 15–25 false starts.
Manual time to find them: 30–60 min.
Sapari time: under 10 min.

You know the shape
"So the way I —"
"Let me start over."
"And if you — wait."
Re-reading a bullet flat.

How it works

Semantic, not
signal-processing.

01 · Transcribe

Word-level transcript

Speech-to-text produces every word with timing.

02 · Detect

LLM reads in chunks

A language model reads overlapping chunks and flags moments where the speaker restarted the same thought.

03 · Score

Confidence threshold

Each detection gets a score. The aggressiveness slider sets the threshold.

This isn't pattern-matching on "um" or "uh" — it's semantic. The model understands that "the way I think — actually, the way I think" is one restart, not two separate sentences.

Controls

Aggressiveness slider.

Position Threshold What it catches
Off
Nothing.
Conservative
0.85
Only the obvious restarts.
Moderate
0.70
Restarts plus clear stumbles.
Aggressive
0.40
Everything the model suspects.

Long false starts (10+ seconds) get a duration penalty so the model doesn't aggressively cut chunks of real content that happen to look like a restart.

Review on the timeline

Purple cards.
You decide.

False start 01:12 · 4.1s

"so the way I — actually, the way I think"

Keep Dismiss
False start 08:34 · 2.8s

"let me — let me start that over"

Keep Dismiss
False start 14:09 · 5.6s

"and if you — wait, sorry, if you want this to"

Keep Dismiss

Why this is rare

Most AI tools only cut silence.

Detecting false starts requires a language model that understands intent — not a signal-processing pass that measures volume.

Sapari runs this as part of the same analysis as silence, captions, audio, and B-roll. One pipeline, one review.

See the full pipeline →

Before you ask

Common questions.

Is it accurate? +

In typical recordings, the Moderate setting catches most true restarts with few false positives. Confidence scoring means you rarely get surprised — obvious restarts score high, ambiguous ones score low and only appear at Aggressive.

What if I want the restart in the final cut? +

Dismiss the card and the restart stays. The AI suggests, you decide.

Does it work in non-English? +

Yes. Transcription supports English, Spanish, Portuguese, French. The LLM handles all four. English is most extensively tested.

Does it catch um and uh? +

Not directly — false start detection focuses on restarted thoughts, not filler. But silence removal at Hyper picks them up because Hyper is tuned to remove anything that isn't speech.

What about comedic or intentional restarts? +

Dismiss the card and the AI leaves it alone. Same review model as every other detection.

23 false starts in 45 minutes.
Found in 4.

7 days. 30 AI minutes. No credit card.

Start free trial