combine-then-transcribe

Combine then transcribe

Stitch together multiple audio files and run them through preprocessing + ASR as a single job. Cheaper (one upload, one job) and produces a coherent single transcript.

When to use

• Several voice memos recorded in one sitting that the user thinks of as one piece
• Recording was split across files due to app/device limits (e.g. WhatsApp 16MB cap, phone storage)
• Multi-part interview where each part is a separate file
• Anything where the user says "combine these and transcribe"

Inputs

• A list of files, a glob, or a folder
• Optional order override (otherwise sort by filename, which usually matches recording order if filenames have timestamps)

Pipeline

1. Validate / harmonise

Before concatenation, all files must share the same sample rate, channel count, and codec. If they don't, normalise each one first via the preprocess-for-transcription skill (or just the format-normalise pass of it). Cheapest: normalise all to mono / 16kHz / opus, then concat.

2. Concatenate

Use ffmpeg's concat demuxer (lossless when codecs match):

# Build the concat list
{
  for f in "${FILES[@]}"; do
    printf "file '%s'\n" "$(realpath "$f")"
  done
} > /tmp/concat-list.txt

ffmpeg -f concat -safe 0 -i /tmp/concat-list.txt -c copy COMBINED.opus

If a boundary marker is desired (helps the user / ASR see file breaks), insert a short silence between segments. Easiest approach: insert a 1s silent opus snippet between every two files in the concat list.

3. Emit segment map

Write <combined-stem>.segments.json capturing where each source file lives in the combined timeline:

{
  "combined": "audio/processed/session.combined.opus",
  "boundary_silence_seconds": 1.0,
  "segments": [
    {"index": 0, "source": "memo-01.opus", "start": 0.0,    "end": 312.4},
    {"index": 1, "source": "memo-02.opus", "start": 313.4,  "end": 654.1},
    {"index": 2, "source": "memo-03.opus", "start": 655.1,  "end": 901.2}
  ]
}

Durations come from ffprobe -i <file> -show_entries format=duration. This sidecar is what makes diarisation timestamps and "where in the recording" references traceable back to the original files.

4. Preprocess + transcribe

Invoke preprocess-for-transcription on the combined file (silence-trim is fine here — boundary silences are 1s, well under the 2.5s threshold, so they'll be preserved as boundary cues). Then invoke transcribe-assemblyai (or whichever ASR the user picked).

5. Optional — split transcript by source

After transcription, offer to split the resulting transcript back into per-source files using the segment map and AssemblyAI's word-level timestamps. Output: transcripts/<source-stem>.from-combined.md for each source.

Folder convention

• Sources stay in place (or audio/raw/ if user is organising)
• Combined file: audio/processed/<session-name>.combined.opus
• Segment map: audio/processed/<session-name>.segments.json
• Transcript: transcripts/<session-name>.combined.raw.md (then iterative-refine takes over)

Output report

Inputs:       3 files, 26m 48s total
Combined:     audio/processed/session.combined.opus (mono 16kHz, 4.1 MB)
Segment map:  audio/processed/session.segments.json
Boundaries:   1.0s silence between each source

Next: send to AssemblyAI? (y/n)

Edge cases

• Mixed codecs / sample rates: must re-encode (lossy concat, but for ASR purposes irrelevant). Warn the user.
• One file is corrupt / unreadable: skip it, log warning, continue with the rest. Don't fail the whole job.
• Total combined length > 4 hours: AssemblyAI handles it but costs add up — flag the estimated cost (~$0.37/hr with diarisation) before submitting.
• User passes a single file: just route directly to preprocess-for-transcription + transcribe; no concat needed.