Transcript Cleaner — clean the speech without cleaning the person out of it

Every transcript arrives dirty: the ums and uhs, the double words, the sentence that starts twice, the missing periods. Cleaning is legitimate — a readable transcript gets used, and a dirty one gets ignored. But cleaning has a failure mode that most tools sail straight into: they make every speaker sound the same. Remove too much and the careful speaker and the chaotic one come out identically smooth, the regional voice gets standardized, and the moment of real emotion reads like a press release. This tool takes the opposite stance. It performs a strictly lexical cleanup — filler tokens, stutters, spacing, punctuation mechanics — and refuses every operation that would touch meaning, vocabulary, or voice. You choose the aggressiveness; the personality is not on the menu either way.

What the tool actually changes

Five operations, all of them mechanical, none of them editorial:

1. Filler token removal. At the standard level, the four hesitation sounds: um, uh, ah, er — matched as whole words, so a name like "Uma" survives. At the aggressive level, six discourse fillers join them: like, you know, kind of, sort of, basically, actually.
2. Stutter collapse. A word immediately repeated ("was was") collapses to one occurrence. The regex is word-boundary anchored, so legitimate repetition across phrasing is safe.
3. Punctuation mechanics. Double commas collapse; spaces before punctuation marks are removed; the letter after a sentence-ending mark is capitalized; a terminal period is added if the text ends without sentence punctuation.
4. Whitespace normalization. Runs of spaces become one; the result is trimmed.
5. Nothing else. No paraphrasing, no reordering, no grammar fixes, no synonym substitution, no sentence merging, no summarization. The output is your input minus a defined list of tokens and spacing errors.

The importance of that fifth item cannot be overstated, because it is what makes the tool safe for voices that standard cleaners damage. A paraphrasing "enhancer" must decide what the speaker meant; a lexical cleaner never gets the chance to decide wrongly.

The two aggressiveness levels

Standard removes only the hesitation sounds — um, uh, ah, er. This is the right level when the transcript represents someone: captions, quotes, testimony, anything where the reader should hear the person. Hesitation sounds are noise in text that was signal in speech; removing them costs nothing audible, because no reader misses an "um" they never heard.

Aggressive adds the discourse fillers: like, you know, kind of, sort of, basically, actually. This level suits working documents — internal summaries, research notes, search indexes — where speed of reading beats fidelity of voice. It is the wrong level for anything quotational, and the reason is linguistic: these six items are not pure noise. "You know" checks comprehension, "like" marks approximation and quotation in many dialects, "actually" signals a correction. Stripping them from a quote changes how the speaker sounds to anyone comparing the quote with the audio — which is exactly the comparison audiences increasingly make.

Choose by destination, not by taste: where a human voice is on the record, standard; where only the information matters, aggressive.

Fillers, and the ones that are secretly content

The hesitation sounds (um, uh, ah, er) are the safe removals — they carry no propositional content in any dialect of English, and their presence in text is purely an artifact of transcription. The discourse fillers are more interesting. Research on spoken language treats them as doing real work: holding the floor, flagging upcoming hesitation, signaling that what follows is approximate. In transcript cleaning, the practical rule is that a filler doing work in the audio is doing work in the text too, even though the text reader cannot hear it.

Three cases where even hesitation sounds deserve a second thought. Testimony and emotion: a speaker saying "I — um — I never told anyone this" carries the hesitation as part of the meaning; stripping it flattens a moment whose difficulty was the point. Comedy timing: a well-placed "uh" is a beat; comedians use them deliberately. Quoted speech inside the transcript: when a speaker is performing someone else's words, those words — hesitations included — belong to the quoted person, not to the cleaning pass. The tool removes what it is told to remove; deciding these cases is the editor's job, which is why the original stays visible beside the cleaned output.

False starts and stutters: what gets caught, what doesn't

The stutter collapse handles the exact-repeat case: "I was was going to say" becomes "I was going to say." This covers the most common transcription artifact, because ASR systems faithfully record word-level restarts. What it does not cover, by design, is the multi-word false start: "I was going to— we ended up going to the second location." The first clause is not a repeat of anything; it is an abandoned thought, and recognizing it as abandoned requires understanding the sentence — an editorial judgment this tool refuses to make, because the cost of being wrong is deleting something the speaker meant.

The honest workflow for multi-word false starts: clean mechanically first, then read the diff with an ear. False starts are usually visible in text once the noise is gone — the abandoned clause hangs there grammatically orphaned — and removing them one by one with a human decision attached is faster and safer than any automated attempt. The tool's contribution is getting the transcript clean enough that the false starts are the only problem left.

The punctuation mechanics, itemized

Small corrections, large readability effects. Capitalization after sentence marks fixes the most common ASR output style — an endless lowercase scroll that reads as one unbroken thought. Removing spaces before punctuation ("word .") restores normal typography. The guaranteed terminal period matters more than it sounds: transcripts frequently end mid-flow, and a text that ends without sentence punctuation reads as damaged rather than finished, which biases readers against using it at all. Double-comma collapse handles the ASR habit of transcribing a pause as a comma plus a comma. None of these operations can damage meaning — they act on whitespace, case, and marks, never on words — which is why they run at both aggressiveness levels without exception.

What the tool refuses to touch

The refusal list is the feature. Dialect forms stay: "y'all," "ain't," "gonna," habitual constructions, regional vocabulary — nothing in the operation list can see them as targets, because they are not fillers and not stutters. Slang stays, for the same structural reason. Profanity stays: cleaning it is a separate editorial decision with consent and audience implications no mechanical pass should make silently. Grammar stays: "me and him went" is not corrected to "he and I went," because grammar is identity as much as syntax, and the reader who compares text with audio will hear the correction as a lie. Word choice stays, sentence order stays, paragraphing is untouched.

The policy in one sentence: the tool will make the speaker easier to read and never make them sound like someone else. A transcript where every guest reads like the same careful corporate author is not clean; it is homogenized, and homogenization is the specific damage that turns readers against transcripts — they stop believing any of it.

The over-cleaning problem

Over-cleaning is the quiet failure mode of transcript work, and it compounds. One cleaned quote reads fine; a page of cleaned quotes from six different people reads like one person wrote all of them — because one system did, at the same settings, with the same filler list. The audience may not name what is wrong, but they feel the sameness, and the feeling converts into a specific doubt: if the words are managed, are the people real?

The damage is worst exactly where authenticity matters most. In testimony, the hesitation is often the evidence of sincerity. In comedy, the filler is the timing. In expert speech, the "well, actually" the cleaner removes was the expert self-correcting — a credibility signal converted into an error by the cleanup. And in any content where the speaker belongs to a dialect community, standardization reads to that community as erasure: the cleaned version is not their voice with the noise off, it is someone else's voice wearing their words.

The antidote is procedural, not technical. Clean at the lowest level the destination allows, read the diff before accepting, and keep the original stored beside the cleaned version permanently. The tool is built around that procedure: both texts stay on screen, the diff is quantified, and nothing is destructive — the cleaned output is a copy, never a replacement.

Aggressiveness by destination

A working matrix for choosing the level, by where the cleaned text is going:

• Captions and subtitles: standard. Viewers hear the audio and read the text simultaneously; every divergence is noticeable. Remove hesitation sounds only.
• Published quotes and show notes: standard, diff-reviewed. Quotes carry the speaker's reputation; aggressive cleaning creates verifiability risk the moment someone checks against audio.
• Blog posts derived from episodes: standard for quoted speech, aggressive acceptable for narration built from the content — and the two should be visibly distinguished anyway.
• Internal research notes and summaries: aggressive. Nobody is quoting these; reading speed is the whole value.
• Search index and AI context: aggressive. Retrieval does not care about voice, and filler tokens dilute keyword signal.
• Legal or compliance records: none. Do not clean. Archive the verbatim transcript; if a readable version is needed, produce it as a clearly labeled derivative and keep both.
• Archival preservation: none. The archive keeps the original; cleaning belongs to the access copy, which is a separate file with a separate name.

The pattern across the matrix: the closer the text stays to being evidence of what someone said, the less cleaning it should receive.

Reading the diff before you accept

The tool shows the cost of its work in two numbers — how many tokens were removed and what fraction of the original that represents — and those numbers are diagnostic. A typical spoken paragraph loses three to eight percent at standard level; if your diff shows fifteen percent or more, stop and inspect, because something unusual is happening. Common causes: a speaker with a genuine hesitation-heavy style (legitimate, but consider standard instead of aggressive); a passage full of quoted or performed speech (hesitations are content there); or an ASR output so noisy that mechanical cleaning cannot finish the job. Common sense applies in the other direction too: a zero-percent diff on a transcript you know contained fillers suggests the paste lost its content somewhere upstream.

Beyond the numbers, read the cleaned text once, aloud if possible. The ear catches what the eye skips: a removed "like" that was doing quotation work ("she was like, no way"), a collapsed repeat that was emphasis ("never, never do that" deserves a second look), a capitalized letter after a period that was actually an abbreviation. Two minutes of reading converts the mechanical pass into an editorial one.

Worked example: one paragraph, both levels

Note for transcript-cleaner: The transcript below is a hypothetical example — an invented speaker used to show the two levels' behavior, not a real person's words.

Original: "Um, so like I was — I was trying to record and, uh, OBS kept, you know, dropping frames? And I was like... wait, maybe the bitrate was too high?"

Standard level removes um and uh and collapses the "I was — I was" stutter's visible half: "So like I was — I was trying to record and, OBS kept, you know, dropping frames? And I was like... wait, maybe the bitrate was too high?" The discourse markers survive; the speaker still sounds like themselves, slightly tidied.

Aggressive level also strips like and you know: the same content, noticeably flatter, faster to read, and wrong for any context where the reader might hear the audio. Notice what neither level touched: the casual "wait," the trailing question structure, the informal register. Those are voice; the tool has no mechanism to reach them. That is the design working as intended, twice.

What the tool cannot do

The honest capability boundary. It cannot remove multi-word false starts, because recognizing an abandoned clause requires understanding sentences. It cannot fix ASR mis-hearings — a wrong word is a wrong word, mechanically indistinguishable from a right one. It cannot add paragraphing or speaker turns; it operates on the text it is given, structure included. It cannot clean audio or fix timing; captions made from cleaned text inherit whatever timing the captioning step assigns. And it cannot judge when cleaning itself is inappropriate — the destination matrix above is that judgment, made by you. Every limit follows from the same principle: operations that require interpreting meaning are operations this tool will never perform, because the day it does, it can start erasing people.

FAQ

Is anything uploaded or stored? No. Cleaning runs entirely in your browser; the text never leaves the tab. For transcripts under NDA or embargo, that is the difference between usable and unusable tooling.

Why is the "keep slang/dialect" control on by default and effectively permanent? Because the cleaning engine has no dialect operations to disable. The control states the policy where users look for it: slang, dialect, and profanity survive every level, by construction rather than by setting.

Will it break timestamps or speaker labels mixed into the text? Labels like "Sarah:" survive — they are not filler tokens. Inline timestamps can be affected at the punctuation level if they use punctuation the mechanics touch; for timed transcripts, clean the text first and merge with timing afterward, or clean section by section away from the timecodes.

Can I clean a transcript in another language? The filler lists are English. Punctuation mechanics (spacing, capitalization after sentence marks, terminal period) apply broadly, but the filler removal will do nothing for non-English hesitation sounds. Treat non-English results as punctuation-only cleaning.

Does aggressive mode ever remove words that matter? It can remove discourse markers that were doing pragmatic work — quotation-marking "like," correction-marking "actually." That is why aggressive is scoped to working documents, and why the diff exists. The tool is honest about what it takes; the destination choice is yours.

Cleaning in the subtitle pipeline

Order of operations matters when the cleaned text becomes captions. The sequence that keeps every stage reversible: clean the transcript first at standard level, review the diff, and only then generate or align captions from the cleaned text. Cleaning after captioning is a trap — filler removal changes token counts, which shifts every downstream alignment, and suddenly the captions and the transcript disagree about what was said. The same rule applies to chapter generation and clip mining: those tools split on sentence punctuation, so a punctuation-mechanics pass before them improves their output, while aggressive cleaning before mining can remove words the ranking uses as evidence. One cleaning pass, done early, at standard level, with the original preserved — the rest of the pipeline inherits the decision instead of re-making it.

The economics: mechanical first, human second

Manual cleaning of a one-hour transcript runs two to four hours for a careful editor; the mechanical pass runs instantly and removes the layer of work that requires no judgment. The division of labor that follows is the real productivity gain: the tool takes the deterministic work (hesitation sounds, stutters, spacing, capitalization), and the editor takes only the judgment work (false starts, filler-as-content decisions, structural problems). Editors report the judgment pass over a mechanically cleaned transcript is faster and more pleasant than a from-scratch pass, because the remaining problems are visible instead of buried — and the total human time drops by more than half even though the human pass still happens. The alternative, cleaning everything by hand, is not higher quality; it is the same final quality arrived at expensively, with more fatigue-induced errors near the end of long sessions.

When the transcript itself is the product

For shows that publish transcripts as content — for accessibility, for search, for readers who prefer text — cleaning standards tighten again, because the transcript is no longer a working document but the deliverable. Three conventions serve that case. Keep standard level regardless of length; published transcripts represent speakers permanently, and aggressive artifacts age badly. Preserve paragraphing that follows topic shifts rather than arbitrary length, because readers navigate long transcripts the same way they navigate chapters. And publish the transcript with a visible note of what was cleaned — one sentence ("Edited for filler and readability; verbatim audio above") converts the transparency problem into a trust feature. Readers who discover undisclosed cleaning feel deceived; readers who are told about cleaning appreciate it, and the difference is entirely in the disclosure.

Team workflow: who cleans, who reviews

Where more than one person touches transcripts, the cleaning step needs ownership like every other editorial step. The pattern that holds up: the producer runs the mechanical pass and reviews the diff numbers, the editor reads the cleaned text for the judgment cases, and the original file stays in the project's storage under a naming convention that marks it as source (episode-ID-verbatim versus episode-ID-cleaned). Two rules prevent the failure modes. Nobody overwrites the verbatim file — the cleaned version is always a sibling, never a replacement. And any quote that leaves the organization gets checked against the verbatim source by the person publishing it, because accountability for a speaker's words attaches to the publisher, not to the tool. The tool is deliberately the least interesting part of this workflow: deterministic, local, and limited, so that the interesting part — the editorial judgment — stays human where it belongs.

Why transcripts arrive dirty in the first place

Understanding the source of the noise makes the cleaning decisions easier. Automatic speech recognition transcribes what it hears, faithfully — and speech genuinely contains hesitation sounds, restarts, and run-on syntax, so a "perfect" ASR output is dirty by definition. Punctuation in ASR output is inferred, not heard: the periods and commas are a model's guess about prosody, which is why they land wrong often enough to need mechanical repair. Speaker turns, hesitations inside words, and overlapping speech each leave characteristic artifacts — merged words, missing tokens, duplicated fragments — and the stutter collapse exists precisely because duplication is the artifact class ASR produces most reliably. The practical takeaway: dirt in a transcript is not an error state to be ashamed of; it is the expected output of honest transcription, and the cleaning pass is a normal stage of the pipeline rather than a rescue operation.

FAQ — a few more

Can the cleaner remove a specific word everywhere, like a verbal tic that is not on the filler list? Not by itself — the token lists are fixed by design, because user-extensible removal lists are how tools start eating real words. For a personal tic, clean mechanically first, then handle the tic with a manual find-and-replace pass where every removal is visible.

Why did aggressive mode leave a "like" in place? The removal is word-boundary anchored and case-insensitive, but it still operates on tokens; a "like" fused into punctuation without spacing, or appearing inside a hyphenated construction, may survive the match. The diff will show it, and manual cleanup takes one second.