Chunking long speech can cut a semantically complete sentence at a boundary

What it is

When long audio is recognized in several pieces, the knife rarely lands on a period. A chunk boundary can fall mid-clause, after a negation, or before a number is finished. Each piece decodes in its own context; the stitched text shows breaks, repeats, or wrong words at the knife. Chunking keeps the length and can destroy a semantically complete sentence.

Why it happens

Fixed-duration windows ignore syntax. Overlapping sliding windows try to align in the overlap and still cut at low energy that is not a sentence end. Each chunk has its own language-model prefix, so the second half cannot see the first half’s subject; pronoun and negation scope are severed. If stitching is only string concatenation, partial words at the knife (rare for a Chinese character, common for an English word or a multi-syllable Chinese word’s acoustics) become a deletion plus an insertion. This is not premature endpoint truncation: both sides of the audio still exist; what is wrong is cut placement and lost context. Semantic endpointing could offer better knives; implementations often slice on wall-clock to beat a cap.

Studying it

Take long dictation with known sentence boundaries, cut at several window lengths and overlaps, and align the stitch against a whole-clip decode. Count knife-in-clause rate and how often negations and numbers are split. Compare “cut at silence” with “cut at a fixed second.” Read news understates spoken subordinate clauses.

Where it stops holding

Captions may wrap at a phrase boundary and viewers can complete them; an executable command split in half may run the first half. Human post-editing can repair the knife; a live system cannot. Large overlap creates repeated words that must be deduplicated, and dedup can delete intentional repeats. If no single span ever exceeds the cap, this chunking mechanism never starts.

Applying it

• Prefer to slice on longer silences or a semantic endpoint, not on the whole second.
• Keep overlap alignment when stitching, and mark uncertain seams in the UI so people can edit.
• Do not execute commands across chunks: if a chunk is incomplete, wait for the next or ask.

Related

• Same group: C7.15.1 A single recognition request usually has a duration or token cap · C7.15.2 Past the cap the system may truncate, chunk, or refuse · C7.15.4 Length caps directly constrain dictation usefulness
• Adjacent: C7.09 Endpoint Detection and End-of-Utterance · C7.03 Types of Recognition Errors
• Search: chunk boundary · sliding window ASR · sentence split