45,000-character chapter on Sume TTS: split evenly costs 2 cents more

A 45,000-character chapter takes three Sume TTS requests because each request is capped at 20,000 characters. Filling two requests to the cap and sending the remaining 5,000 costs $2.14. Splitting evenly into three 15,000-character requests costs $2.16. The difference is only two cents, but it shows how per-job rounding works.

The arithmetic

Billable price is 47.5 micro-dollars per character, rounded up to a whole cent per job. A full 20,000-character job is 950,000 micro-dollars, exactly 95 cents, so there is no rounding loss. A 5,000-character job is 237,500 micro-dollars, 23.75 cents, which rounds up to 24. A 15,000-character job is 712,500 micro-dollars, 71.25 cents, which rounds up to 72.

Rounding loss is per job, so fewer jobs lose less

Each job can lose up to just under one cent to rounding. Three jobs lose at most about three cents in total, and the even split lost 0.75 cents on each of three jobs. That is why filling the cap is slightly cheaper, but it is not a reason to cut in the middle of a sentence.

Cut at paragraph or scene breaks instead. A hard cut inside a sentence changes the prosody at the seam, and no price saving of two cents is worth a retake. If the break falls at 17,300 characters, send 17,300: 17,300 x 47.5 = 821,750 micro-dollars, 82.175 cents, billed 83 cents.

When even splits make sense

Sometimes the even split is the better engineering choice. Three jobs of about 15,000 characters finish at similar times, and a failed job costs you less to retry. Idempotency keys are required on writes, so give each chunk its own key (for example chapter-07-part-1) and a retry will not double-spend.

The fill-to-cap plan has a worse failure mode: if the first 20,000-character job needs a retake, you redo 95 cents rather than 72. At these prices the choice is about workflow, not savings.

Joining the files

Timeline audio concat joins up to 20 Sume-hosted audio parts into one gapless file. The join is in the sample domain with no re-synthesis and no silence at the seams, and the result lists the offsets of each segment. Import any outside audio into the workspace first, since the server only reads media.sume.com files from your workspace.

The concat itself adds no gaps, so write any pause between chapters into the script text.

Checking the result

After the three jobs finish, confirm each audio artifact's duration against the script, then concat in order. If one chunk sounds different, read the job record for model_id, voice, language and speed and compare it with the other two; a mismatch in any of those is the usual cause. A retake of a 15,000-character chunk costs 72 cents; of a 5,000-character chunk, 24 cents.

Chunk by scene or heading, name each chunk's Idempotency-Key after it, and keep a table of characters per chunk so you can predict the bill before you submit.