Segment joining: low-energy cuts, ASCII spaces and original-word evidence

Joining may add a boundary space; it does not certify grammar, punctuation or meaning

CollieNoteSegmenter uses 16 kHz, 28-second windows and a low-energy search over the last eight seconds in 1600-frame steps, with final/short-window handling. join inserts one space only when both adjacent characters are ASCII letters/digits; other boundaries concatenate. It is not semantic splitting, punctuation repair, diarization or word alignment. No structured-minutes engine is added.

Different from original-word and encoding review

Listening reviews content; export review checks serialization. Here what segmentation/joining itself adds or preserves at boundaries matters, so processed text is not mistaken for an untouched verbatim original. No model/audio benchmark ran.

Check Verified scope Manual review Not established
Window 16 kHz/up to 28 seconds Final window/actual length Each window is a whole sentence
Search Last eight seconds, low energy Acoustic versus semantic cut Words never cut
Step 1600 frames, about 100 ms Granularity versus content 100-ms word alignment
ASCII boundary Letters/digits both sides add space API + V2 → API V2 Space necessarily came from speech
Other boundaries Direct join, no rhetoric repair CJK/punctuation/existing spaces Grammar correction/paragraph hierarchy
Empty/first piece Existing/next per implementation Original words/output Missing sentences generated
Progress Frame position differs from text transcribedFrames/audio Every piece manually reviewed
Retention Original/processed/edited text separated Added spaces/editorial reasons Automatic verbatim authenticity certificate

Synthetic joining illustration

API + V2 becomes API V2; 样例 + 说明 becomes 样例说明. These illustrate boundary rules, not recognition accuracy. Manual punctuation, splitting or alias edits must retain editorial status/evidence rather than claim the model/original explicitly supplied them.

Low energy is not guaranteed word absence. Review actual audio/text. Later briefs still preserve conditions, negation and unknowns; joining does not establish business approval or execution.

FAQ

Are 28-second windows semantic segmentation?

No. Acoustic windows/cuts differ from sentence boundaries or speaker separation.

Do low-energy cuts guarantee no split words?

No. Strategy does not replace actual audio/result review.

Must the API V2 space come from speech?

No. join can insert it at ASCII letter/digit boundaries. Distinguish processing from originals.

Does CJK joining add spaces/punctuation?

Not through this join. Other boundaries concatenate; it is not general typography/correction.

Does frame progress establish all words heard/reviewed?

No. Position, final text and manual acceptance differ.

Is new segment editing or automatic minutes added?

No. This reviews fixed implementation, not new UI/API/model capability.

Reading and source

Original-word review · Export format · Audio/text mapping · Start · Privacy

Segmenter/join · Frame progress metadata