The raw output problem
Anyone who has looked at an unprocessed machine transcript knows the look: a long run of words, perhaps all lowercase, with no commas, no full stops and no clear paragraphs. The words may be mostly right, yet the text is tiring to read and easy to misread. The reason lies in how speech works and how recognition software is built.
Speech does not contain punctuation
Writing and speech are different systems. In writing, punctuation marks show how sentences are built. In speech, nothing equivalent is spoken aloud. Speakers use pauses, pitch, rhythm, emphasis and breath, and listeners combine those cues with grammar and context to work out where one thought ends and the next begins.
The basic job of recognition is to decide which words were said. Punctuation and capital letters are not sounds, so they do not come directly from the audio the way words do. Many systems therefore first produce an unpunctuated string of words, and a separate step adds structure afterward.
How tools add punctuation
Automatic punctuation is typically a second guess layered on the first. A model looks at the recognized words, and often at some timing or pause information, and predicts where sentence boundaries, commas, question marks and capital letters are likely to belong. Some modern systems predict punctuation as part of recognition itself. In either case it is an estimate, based on patterns learned from large amounts of text and speech.
The estimate is decent for clear, well-paced speech that follows ordinary sentence patterns. It tends to struggle in other cases.
Why it goes wrong
Pauses do not equal sentence ends. People pause to think, breathe or search for a word in the middle of a sentence, and they sometimes run two sentences together without a break. A tool that relies heavily on pauses may insert a full stop mid-thought or miss a real boundary.
Spoken language is untidy. Natural speech contains false starts, repeated words, interrupted sentences and trailing off. There may be no single correct punctuation for something that was never a tidy sentence.
Questions without question words. A rising pitch can turn a statement into a question (“You’re coming tonight?”). If the tool relies on the words alone, it may miss this.
Commas are judgment calls. Even in writing, comma use varies by style. Automatic systems may add too many, too few, or put them in odd places.
Casing of names and terms. Capital letters at the start of sentences depend on sentence boundaries being right. Capitals for proper nouns depend on recognizing that a word is a name, which is hard for unfamiliar names and brand terms. Acronyms, product names with unusual capitalization, and words that are both common and proper nouns all cause trouble. Sometimes the tool capitalizes too much, such as after a mistaken full stop.
Numbers, symbols and formatting. Tools may write numbers as words in one place and as digits in another, and may or may not add symbols such as currency signs or percentages. Consistency within one transcript is often poor, which is where number formatting rules come into play for subtitles.
Other languages and mixed speech. Punctuation conventions differ between languages, and systems typically work better for widely used languages. When a speaker switches languages mid-sentence, structure guesses become less reliable.
Why this matters
Punctuation changes meaning. The classic example of a missing comma changing an instruction shows how a small mark can make a large difference. A misplaced full stop can split a sentence so that the second half seems to refer to something else. For people reading captions, poor structure also slows reading and increases effort.
For subtitles specifically, the needs differ somewhat from a plain transcript. Subtitle punctuation follows its own conventions, partly because lines are short and time on screen is limited, and breaks in captions are decided by line length and timing as well as grammar. A machine-punctuated transcript is a starting point for these conventions, not the end product.
Fixing it efficiently
A practical order of work is:
- Fix wrong words first. Punctuation around a wrong word is wasted effort.
- Set sentence boundaries. Listen to where thoughts actually end and correct the full stops and capital letters. This single step improves readability the most.
- Add or repair question marks by listening for questions that the words alone do not reveal.
- Adjust commas for clarity rather than chasing a perfect style.
- Check proper nouns and acronyms. Capitalize names and terms consistently, ideally using a glossary or a custom vocabulary for repeat material.
- Standardize numbers and symbols according to the project’s style.
- Add paragraph breaks in a transcript, or reshape lines into cues for subtitles.
A find-and-replace for repeated patterns can help, but be careful with automated cleanup scripts, which can introduce their own errors by changing things that were already right.
What to expect from tools
Some tools let you switch automatic punctuation on or off. It is worth testing both on your own material. For some content, such as a very noisy recording, the punctuation may be so unreliable that you prefer to add it yourself. For clean, well-paced speech, the automatic version may save real time.
Expect improvement over time, but not perfection. A human pass remains necessary wherever readers depend on the text being accurate, and punctuation is part of accuracy, not decoration.