400 NPC voice lines for an indie RPG: $2.00 on Sume, batched

A 400-line NPC voice pass at about 90 characters per line costs $2.00 on Sume if you send 20 lines per TTS job and split the result with timeline audio, against $4.00 if you send each line as its own job. The gap comes from rounding, not from the voice: a single 90-character line is worth under half a cent, but each job is billed in whole cents (read 2026-10-03).

This is a budget for a small studio that wants a placeholder-quality voice pass in the build before the year-end push, roughly 90 days from today. It does not replace casting for your main characters. It covers barks, shopkeepers, quest-givers and the long tail of lines nobody will cast.

The arithmetic

Sume text to speech is $0.0475 per 1,000 characters, spaces and punctuation included, with up to 20,000 characters per request. Twenty lines of 90 characters is 1,800 characters, which is $0.0855 and bills as $0.09. Twenty such jobs cover 400 lines for $1.80.

A batched job returns one audio file, so you need the lines back as separate files. Timeline audio can slice one hosted file into up to 20 ranges per job at $0.01. One split job per TTS job adds $0.20 across the run. If you request word timestamps with sentence segmentation, the TTS result gives you the sentence start and end times to use as the ranges, as long as each line is one sentence. Multi-sentence lines need you to pick the boundary yourself.

What to watch

A single TTS request that would run past 1,200 seconds of audio fails, so 20 lines per job is a comfortable size; check the audio length of your longest batch before you raise it. Keep the same voice id across all batches for one character, and write one script file per character so a rewrite of one line only re-runs that character's batch.

Character and place names are where cheap voice passes break. The TTS request accepts a pronunciation dictionary id, so build one dictionary per game and reuse it in every job. Listen to the first batch before you run the other 19; a mispronounced town name multiplied by 400 lines is the expensive mistake here, not the cents.

Rewrites cost the same as first passes. If 40 lines change after a script edit, re-run only the batches that contain them (at most two batches if the edits cluster) rather than the whole set. Send a fresh idempotency key for each intentional re-run, since a repeated key replays the earlier result instead of generating again.