Transcribing Long Videos by Segment: Practical Method

The two problems with long video transcription

Long videos run into two distinct problems: file size and accuracy. Most transcription services cap direct uploads at 500MB or 90-120 minutes. A 90-minute conference recording at 1080p is typically 2-4GB — three to eight times over the limit. The file never even reaches the transcription engine.

The second problem is accuracy drift. Speaker fatigue causes pronunciation shifts after 30-40 minutes of continuous speaking. Background noise patterns change. Transcription models trained on shorter controlled audio segments handle these shifts poorly when they appear at minute 70 of a continuous stream. Segmenting gives each chunk a clean start. When I processed a 3-hour conference recording as a single file, accuracy noticeably declined in the final hour compared to earlier sections. Processing the same content as three separate one-hour segments produced consistently better results throughout.

Before splitting: reduce file size first

Transcription only uses audio. Video resolution adds nothing. A 90-minute talk at 1080p is 2-3GB. The same talk as an extracted MP3 at 128kbps is approximately 85MB. Extract audio first, before considering any split.

FFmpeg is the fastest tool for audio extraction. One command:

ffmpeg -i recording.mp4 -vn -acodec mp3 -ab 128k audio.mp3

If the resulting audio file still exceeds the service limit, then split. For most 60-90 minute videos, audio extraction alone brings the file well within limits. The FFmpeg documentation covers every conversion option in detail if you need formats beyond the standard MP3 extraction.

How to split at the right points

Splitting mid-sentence causes two problems: the first segment ends with an incomplete sentence (affecting readability and the model's sentence boundary detection), and the second segment starts mid-thought (context the model does not have). Split at natural pauses: between talks in a conference, between chapters in a course, at Q&A transitions, or at minimum between paragraphs.

To find the right split points without rewatching the full video: open the video in any player at roughly the halfway mark. Scrub a few seconds to find a pause. Note the timestamp. Use FFmpeg to split at that point:

# Split at 32 minutes 15 seconds
ffmpeg -i audio.mp3 -to 00:32:15 segment1.mp3
ffmpeg -i audio.mp3 -ss 00:32:15 segment2.mp3

Reassembling the transcript

After transcribing each segment separately, reassembly depends on the output format. For TXT: paste segments in order, add a section break between them if needed. No timestamp adjustment required.

For SRT files: timestamps in each segment start from 00:00:00,000 (the segment start). When combining into one file, you must offset each segment's timestamps by the cumulative duration of all preceding segments. If segment 1 runs to 32:15, every timestamp in segment 2 increases by 32 minutes and 15 seconds. Sequence numbers also need to be renumbered sequentially across the full combined file.

The guide to transcript segmentation and timestamps covers the timestamp offset logic in detail, including the SRT and VTT format differences for multi-segment files.

Practical workflows for common long-video scenarios

Conference recordings are the most common long-video transcription use case. A one-day conference with eight 45-minute talks is best handled as eight separate transcription jobs, one per session. Each transcript gets labelled with the speaker name and session title before storing. When you need to review the event later, eight individually searchable text files are more useful than one 8-hour transcript that requires scrolling to locate any specific session.

University lectures follow the same logic. A semester of lecture recordings, transcribed as individual weekly files named by topic, creates a searchable study archive. Searching the folder for a specific term locates every lecture where that concept was mentioned. That's not possible with video files and barely possible with a single merged transcript that has no section labels.

Podcast episodes with remote guests present a specific challenge: two speakers, often with different microphone qualities. If the host is on a studio microphone and the guest is calling from a phone, the mixed output creates uneven audio that affects accuracy. Segmenting at question-and-answer boundaries and processing guest responses separately from host questions can measurably improve the result. Test it with one episode to decide whether the extra step is worth it for your specific setup.

The host-guest audio quality gap is the most common problem in podcast transcription. Host on a studio mic. Guest on a phone. Two different accuracy levels in one file. Segment at the guest boundary. Process separately. Merge after review.

When processing multi-speaker content where audio quality varies significantly between participants — a common scenario in podcasts with remote guests — segmenting at speaker boundaries and processing each segment separately produces better results. The overhead is real (15–20 minutes per episode of additional workflow), so it is only worth applying when the accuracy difference between speakers is significant and the final output needs to be publication-quality.

Verifying quality across segments before finalising

Segment boundaries are the highest-risk locations for errors: the last sentence of one segment and the first sentence of the next can lose context when the model starts fresh. A sentence split mid-clause by a boundary sometimes produces confusion about what preceded it. A targeted quality check covers: the first paragraph, the last paragraph, the three or four segment boundary points, and any passage where the speaker's audio was noticeably unclear. Catching errors at specific locations is more efficient than a full word-by-word pass before reassembly.

For each segment, a targeted five-minute quality check covers the essentials: read the first paragraph, read the last paragraph, spot-check three sentences in the middle, note errors on a list. Batch corrections at the end are faster than fixing inline as you go.

For high-accuracy use cases — legal transcripts, published journalism, academic citation — full review is required before reassembly. Segment boundaries are where errors compound: a sentence split mid-clause at a segment boundary loses the context that preceded it, and the model processing the second segment doesn't know what came before. Always verify the first and last sentence of every segment join manually.

The transcript proofreading guide covers the full review process with time estimates per hour of source audio, useful for planning how long the segment review stage will take on a long recording.