Fill in the safeguard that keeps words from vanishing at audio chunk boundaries
Overlapping chunks of audio (typically 2 to 5 seconds shared between neighbors) prevent words sitting on a boundary from being clipped, and word-level timestamps let you merge the duplicates back out cleanly.
Imagine cutting a movie into 30-second clips and watching them out of order. If you cut exactly at the moment someone says the word important, half the word ends up on one clip and half on the next. Whoever transcribes those clips may write impor on one and tant on the other, missing the word entirely. The fix is to let each clip start a few seconds before the previous one ended, so the full word lives somewhere intact in at least one clip. Later you compare timestamps and throw away the duplicate seconds, keeping a clean transcript with no words cut in half.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Long-form speech recognition is mostly a chunking problem. The model itself only takes a window of audio at a time, so anything longer has to be cut, processed in parallel, and stitched. The interesting question is how to cut without losing the words at the seams.
This deep dive covers why naive cutting fails, why overlap solves it, and why the same shape appears across every long-content pipeline in ML.
Why fixed cuts lose words
Speech models decode each frame of audio in the context of the surrounding frames. A word at position 30.0 seconds in a chunk that ends at 30.0 seconds has zero right-context: the model sees the start of the word and nothing after it. Acoustically similar words become indistinguishable (cat vs. cap, three vs. tree), and the decoder either picks the wrong one, drops the word, or hallucinates a plausible-sounding token.
This is not a model bug. The same word in the middle of a chunk decodes correctly because its surrounding phonemes constrain the search. The bug is in the chunking strategy: it forces some words to live without their context.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- OpenAI Whisper's reference implementation uses a 30-second window with a small overlap and timestamp-based stitching for long-form audio
- Deepgram Nova-3 exposes word-level timestamps explicitly so callers can implement overlap stitching
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you pick the overlap length for a new language or codec?
Start at 2 to 3 seconds (covers most spoken words) and measure WER on a labeled set. If boundary errors persist, increase overlap by 1 second at a time until WER stops improving. Compression and reverb in some codecs need more right-context, so the right number is empirical.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Chunking at fixed 30-second marks with no overlap, then wondering why every fifth utterance is missing the first or last word of a sentence.
60 second bullets to scan on the way to the call.
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.