47 voiceover lines, Timeline's 20-part limit: three concats, 3 cents

Joining 47 voiceover lines into one Timeline spine takes three concat jobs, 20 + 20 + 7 parts, for $0.03 in total, because Timeline audio accepts at most 20 parts per concat. The render can then take the three joined files as its own audio.parts, which also allows up to 20 slices and adds no job.

The numbers

Each concat is $0.01 flat per job (Timeline audio). The join is sample-domain, with no re-synthesis and no silence at the seams, and every part must have the same channel layout. The result returns segments[] with start offsets that you use to re-base Timeline video[].start. The other number to notice is the TTS floor: 47 lines of 140 characters list at $0.0067 each, but the catalog minimum is 1 cent per job, so 47 jobs bill 47 cents, not the 31 cents that character math on 6,580 characters would suggest.

Computing the groups

This script prints the grouping and the costs. It makes no network calls.

lines = 47
groups = []
while lines > 0:
    n = min(20, lines)
    groups.append(n)
    lines -= n
print("concat jobs:", len(groups), groups)
print("concat cost: $%.2f" % (0.01 * len(groups)))
print("tts floor cost: $%.2f" % (0.01 * 47))
print("render audio.parts:", len(groups))

Choosing between per-line files and one file

Per-line files make a later fix cheap (re-synthesize one line, then re-concat), but they pay the 1-cent floor on every short line. One file is cheaper to generate and can be cut back into ranges with the split operation ($0.01 per job, up to 20 ranges), but a changed line means regenerating the whole script. At 47 lines, per-line files cost 47 cents of TTS plus 3 cents of concat, 50 cents in all, against about 31 cents for one 6,580-character job: roughly 19 cents more up front in exchange for single-line edits.

Edge cases the docs name

Concat needs at least one part and refuses a top-level url or ranges field with audio_concat_takes_no_url and audio_concat_takes_no_ranges, so a request for the third group of seven lines must still use parts[] and nothing else. The produced audio is limited to 1,800 seconds, which is also the longest a Timeline render can carry, so 47 lines that add up to more than 30 minutes need a different plan. Each part can carry a source_in and a duration if a line has a few hundred milliseconds of breath at the start or end that you want to trim without re-synthesis.

If you place the three joined files in the render as audio.parts, remember that Timeline's own parts list is also capped at 20 slices and is sample-domain too, so the whole chain from 47 lines to the final render never re-synthesizes anything.