Voice 10,000 product descriptions: $142.50 on paper, $200 per request

A shop owner wants an audio line for every product page: 10,000 descriptions of about 300 characters each. The first calculation takes ten seconds. The second one, about how requests are rounded, changes the budget by a third.

The paper price

Sume TTS 1.0 costs $0.0475 per 1,000 characters. Ten thousand descriptions of 300 characters is 3,000,000 characters, and 3,000 x $0.0475 = $142.50.

The rounding

Each request's quote rounds up to a whole cent. A 300-character description costs 300 x $0.0000475 = 1.425 cents, which quotes at 2 cents. Ten thousand requests are $200.00, which is $57.50 above the paper price. That is the cost of one description per request. The ceiling bites hardest when requests are short.

What batching costs you

Joining descriptions into one transcript gives you one audio file per request. To get one clip per product you must cut it. Ask for timestamps: {"words": true} plus segmentation: {"mode": "sentence"}. With output_format wav or raw, each segment can carry its own audio_url. With mp3 you only get segment timings, and you cut the file yourself, as the segmentation post explains. Segments are sentences, so write each description as one sentence, or map several sentences back to a product by your own index. Check the cut points on 20 products before you run the lot.

Queue and time

The admission rules cap how many jobs run at once and how many may wait: Pro runs 4 and queues 20, Scale runs 20 and queues 100. With 500 requests, submit in waves of the queue capacity and honor retry-after on a 429. With 10,000 single requests you would spend more of your run waiting on admission than on audio.

Pick by your constraint

If product pages are the product, the extra $57.50 for one-per-request is cheap insurance against bad cuts. If the cost per page matters more, use batches of four to twenty. Run a 100-description pilot at both settings. Compare the bill from the pilot against the table, not your guess.