Transcribing long recordings: split them properly instead of hoping
August 19, 2026 · 8 min read
Two-hour lectures and full conference days transcribe far more reliably when you cut them in the right places first. Where to split, how to fix timestamps afterwards, and the commands that do the work.
A 90-minute lecture, a conference day, a long narrative interview: recordings like these are normal as soon as you transcribe regularly. Sending one through a model in a single piece looks convenient — until a failure at minute 80 costs you the whole file, the transcript turns into an unreadable wall, or the daily allowance is gone. Splitting first is faster, and a mistake costs you a few minutes instead of two hours.
When is splitting actually worth it?
Not every long file needs to be cut up. Three questions decide it: does the file fit the service limits, do you need consistent speaker labels across the whole length, and how painful would a repeat run be? At Sprechverlauf the free plan covers files up to 30 minutes, 90 minutes of audio per day and 10 hours per month, so longer recordings have to be split anyway.
| Recording length | Recommendation | Reason |
|---|---|---|
| Up to 30 minutes | Leave it in one piece | Fits one file and keeps speaker labels consistent throughout |
| 30 to 60 minutes | Two parts, cut at a pause | One part stays repeatable without reprocessing everything |
| 1 to 3 hours | Sections of 20 to 30 minutes | Matches agenda items, breaks, or topic changes |
| Full conference days | One file per talk or session | Every section gets its own sensibly named transcript |
| Recordings with long pauses | Remove the silence first | Saves processing time and avoids empty stretches |
The one rule: never cut mid-sentence
Speech recognition works from context. Split a sentence down the middle and both halves lose it: part one trails off, part two starts with no reference. This hurts most with technical terms and proper names, exactly where the model leans on surrounding words. Cut during a pause instead — ideally at a point where the topic changes anyway.
- Cut between agenda items, slides, or question rounds
- Leave two to three seconds of overlap so no word disappears at the seam
- Drop the greeting, the audio check, and the goodbye entirely
- Remove pure music, applause, and setup passages beforehand
- Name parts in order: “lecture-01”, “lecture-02”, not “new” and “new-final”
Splitting with ffmpeg: two commands are enough
ffmpeg is free, runs everywhere, and will not re-encode your audio unless you ask it to. This command cuts a file into 25-minute sections (1500 seconds) and leaves the audio untouched:
- ffmpeg -i lecture.m4a -f segment -segment_time 1500 -c copy part_%02d.m4a
- ffmpeg -i lecture.wav -ac 1 -b:a 64k lecture.mp3 (shrinks speech recordings considerably before splitting)
- ffmpeg -i lecture.m4a -ss 00:12:30 -to 00:41:00 -c copy excerpt.m4a (cuts out one specific section)
Mono at 64 kbit/s sounds drastic but is uncritical for speech: the model needs neither stereo width nor music-grade resolution. An hour of talking fits comfortably into roughly 30 MB. Keep the original and work from the smaller copy — you always split the copy, never the master.
Splitting without the command line
If you would rather not open a terminal, Audacity gets you just as far: open the file, select a range, then “File → Export → Export Selected Audio”. The waveform view has a genuine advantage over the command line here — pauses are visible as narrow sections, so you can see the cut point instead of guessing it. On a phone, the built-in voice memo apps do the same job, if more roughly.
After splitting: timestamps restart at zero
Every transcript counts from the start of its own file. Part two of a lecture therefore begins at 00:00 again, even though it really starts at minute 25. When you merge the parts, you have to add the start offset to every timestamp, or all your references point to the wrong place. For SRT and VTT files, most subtitle editors do this under a function called “shift” or “offset times”.
- Write down the start offset of each part before you lose sight of the original file
- Add a subheading per part when merging text exports
- Cite quotes with part number and time, for example “part 2, 07:41”
- Merge subtitles only after the times have been shifted
Speaker labels do not carry across parts
Speaker detection analyses each file on its own. “Speaker 1” in part one and “Speaker 1” in part two need not be the same person — neither run knows about the other. That is a property of the method, not a bug. In practice: listen to the first seconds of each part and replace the generic labels with real names before merging the sections.
When consistent speaker attribution matters more than anything else — a hearing, say, or a group discussion with six participants — there is a good case for not splitting at all, setting the speaker count explicitly and accepting the extra processing time.
A workable order of operations
- Archive the original and create a working copy as mono MP3
- Skim the recording once and mark cut points at the pauses
- Split into 20 to 30 minute sections, numbered in order
- Transcribe sections one by one, setting language and key terms first
- Replace speaker labels with real names per section
- Shift timestamps by the start offset and merge the texts
- Spot-check numbers, names, and quotes against the audio