Why does automatic punctuation get worse on phone audio?

Narrow-band phone audio loses the pitch detail the model uses to detect sentence endings, and overlapping speech removes the pauses entirely. Expect more run-on sentences and fewer question marks.

Should I punctuate a full verbatim transcript differently?

Yes. Full verbatim keeps false starts and repetitions, so punctuation marks the breaks as they actually happened, including trailing dashes for interrupted speech, rather than tidying them into complete sentences.

Does punctuation affect subtitle line breaks?

It does. Cue splits should fall at clause boundaries, so incorrect punctuation leads directly to awkward line breaks when the transcript is converted to captions.

Can I turn automatic punctuation off?

Some systems allow it, and unpunctuated output is occasionally preferred where a downstream tool applies its own rules. For human reading it is almost always worse.

Definition

What Is Punctuation Restoration?

VideoText editorial team. Updated September 13, 2026.

Direct definition

Punctuation restoration is the step that adds periods, commas, question marks, and often capitalization to raw ASR output that was predicted as an unpunctuated word stream.

The sections below explain how punctuation restoration is used in transcription and subtitle work, what people often mix up, and which VideoText workflow applies when the next step is a file or review pass.

Key takeaways

Word error rate can look fine while punctuation is unusable.
Question vs statement errors change meaning.
This is an NLP post-process, not a microphone setting.

Why it matters

Captions that lack sentence boundaries become reading-speed problems. Transcripts without punctuation fail QA even if every lexical word is right.

How the model decides where punctuation goes

Speech does not contain punctuation. A restoration model infers it from prosody — pauses, pitch movement, the falling intonation that ends a statement — and from the words themselves, since some sequences are far more likely to end a sentence than others.

This is why punctuation quality varies with audio quality. A clean studio recording gives the model the pauses it needs; a noisy conference call with people talking over each other gives it far less to work from, and the output tends toward long run-on sentences.

What to check in review

Sentence boundaries first: a misplaced full stop changes where a quote begins and ends, which matters more than a missing comma. Then check question marks, which are frequently dropped when a speaker uses statement intonation to ask something.

Do not over-correct. Adding commas to match written style guides makes the transcript read well but moves it away from what was said, which is the wrong trade for verbatim work.

Frequently asked questions

Why does automatic punctuation get worse on phone audio?: Narrow-band phone audio loses the pitch detail the model uses to detect sentence endings, and overlapping speech removes the pauses entirely. Expect more run-on sentences and fewer question marks.
Should I punctuate a full verbatim transcript differently?: Yes. Full verbatim keeps false starts and repetitions, so punctuation marks the breaks as they actually happened, including trailing dashes for interrupted speech, rather than tidying them into complete sentences.
Does punctuation affect subtitle line breaks?: It does. Cue splits should fall at clause boundaries, so incorrect punctuation leads directly to awkward line breaks when the transcript is converted to captions.
Can I turn automatic punctuation off?: Some systems allow it, and unpunctuated output is occasionally preferred where a downstream tool applies its own rules. For human reading it is almost always worse.

Related terms

Sources

OpenAI: Whisper paper (text normalizer discussion)