Transcript cleanup

Clean speech-to-text artifacts from transcripts and direct quotes without editorializing. The goal is readability, not rewording. Preserve the speaker's vocabulary, sentence structure, and personality. Fix only what a competent transcriptionist would fix.

The rule

If you cannot point to a specific speech artifact or transcription error, do not change the sentence. "Sounds better" is not a reason to edit. Every edit must fall into one of the pattern categories below.

Pattern categories

1. Filler "like"

Spoken English uses "like" as a verbal pause. Remove it when it serves no grammatical function. Keep it when it means "such as" or "similar to."

Before	After	Why
has like a road map	has a road map	filler before article
is like no more than three digits	is no more than three digits	filler before adverb
was like AI is important	where the message was AI is important	filler replacing a clause
a mega corp like Apple	a mega corp like Apple	means "such as" -- keep
a generic topic like generative engine optimization	a generic topic like generative engine optimization	means "such as" -- keep

Test: Remove "like" and read the sentence. If it still makes grammatical sense and the meaning is unchanged, the "like" was filler.

2. Spoken grammar

Speakers routinely break grammar rules that readers notice on the page. Fix subject-verb agreement and comparison constructions.

Before	After	Rule
there's so many signals	there are so many signals	subject-verb agreement ("signals" is plural)
as powerful than if we had	as powerful as if we had	"as...as" comparison, not "as...than"
as much of a signal than if	as much of a signal as if	same pattern

Test: Read the sentence aloud slowly. If the grammar error is obvious when spoken deliberately rather than quickly, fix it.

3. Word order from natural speech

Speakers front-load or rearrange words in ways that read awkwardly on paper. Restore standard English word order.

Before	After
put in specifically keywords	specifically put in keywords
that's like deep on a bottom of funnel	that's deep on a bottom-of-funnel

4. Orphan words

Mid-sentence restructuring during speech leaves behind words that no longer connect to anything. Remove them.

Before	After	Orphan
things that while seem exciting	things that seem exciting	"while" left over from an abandoned clause

5. Transcription artifacts

Speech-to-text engines sometimes drop words, merge sentences, or mishear connectives. Restore the minimal missing words needed for the sentence to parse.

Before	After	Fix
was like AI is important	where the message was AI is important	restored dropped clause
who is higher up in the organization chart like a chief marketing officer	who is higher up in the organization chart, like a chief marketing officer	added comma before "like" (parenthetical example)

Only add words when the sentence is genuinely unparseable without them. If the meaning is clear despite missing words, leave it alone.

Process

1. Read the full transcript or quote set before editing anything.
2. Identify each artifact by pattern category (1-5 above).
3. Make the smallest edit that fixes the artifact.
4. Re-read the edited version aloud. If the speaker's voice is gone, revert.
5. When in doubt, leave it. A slightly rough quote is better than a polished one that no longer sounds like the person who said it.

What not to touch

• Informal vocabulary ("gonna", "kind of", "a lot") -- these are voice, not errors
• Sentence fragments that work as emphasis -- speakers use these deliberately
• Repeated words used for emphasis ("it's really, really important")
• Regional or personal speech patterns -- do not normalize dialect
• Content or meaning -- never change what the speaker said, only how clearly the transcription conveys it
• Sentences that are merely inelegant -- ugly-but-clear is fine

Working with direct quotes

When cleaning quotes embedded in documents (blockquotes, inline quotes), apply extra caution. The reader knows this is a quote from a real person. Overcleaning makes quoted speech sound ghostwritten.

For attributed quotes (with a speaker name), preserve more roughness. The attribution signals "this is how they actually talk." For unattributed quotes used as pull-quotes or callouts, slightly more cleanup is acceptable since the reader expects polished text.

Batch editing workflow

When cleaning an entire transcript or document with many quotes:

1. Grep for all direct quotes across the file set.
2. Categorize each fix by pattern (1-5) before editing.
3. Apply fixes file by file, not pattern by pattern, to maintain reading context.
4. Flag any borderline cases for human review rather than guessing.