Chunking — How Long Audio Gets Transcribed

Why chunking exists

Transformer models scale quadratically with input length — doubling the audio length quadruples compute and memory. Practical neural ASR runs on fixed-size windows (Whisper: 30 seconds; some models: 60-90 seconds). For longer audio, you have two options: a streaming model that processes a sliding window, or chunking the input into pieces, transcribing each, and concatenating. We use the second approach with 10-minute chunks per parallel worker.

Why chunk boundaries matter

Naive chunking at fixed-second boundaries can split a word or sentence in half — "transcrip-" + "-tion" instead of "transcription". To prevent this, chunkers use VAD (voice activity detection) to find silence boundaries (>200 ms of silence) and chunk only on those. Result: every chunk starts and ends in silence, no words split, no boundary artifacts in the final transcript.

How we chunk in production

When you upload a file longer than 10 minutes, our pipeline runs Silero VAD to find candidate split points, picks split points near each 10-minute mark that fall inside silence regions, and sends each chunk to Whisper in parallel. Timestamps are normalized — chunk 2's internal 0:00 becomes 10:00 in the merged transcript. The user sees a single seamless transcript, not five glued-together chunks.

In practice

A 2-hour interview becomes 12 chunks of ~10 minutes each. Each chunk takes ~12 seconds to process on Whisper. Run all 12 in parallel: total wall-clock time is the slowest chunk, around 15 seconds — instead of 144 seconds serial. The user gets the full transcript almost as fast as a 10-minute file.

Related terms

• Whisper
• VAD
• Timestamp

Last updated: September 20, 2026