How this TTS calculator estimates batch time
Start with the audio you want to keep. Multiply that duration by extra takes and the time your complete rig needs to generate one minute of audio. Example inputs: 10 finished hours, 50% extra takes and 0.5 minutes of generation per audio minute need 7.5 rig-hours.
Monthly capacity uses the hours you allocate minus reserved idle time. The completion window counts the generating hours available before your deadline. Exclude breaks and other jobs from that window. A monthly budget can fit while its deadline does not. GPU-hours count every GPU in the tested rig; the calculator does not assume that adding cards speeds it up.
Measure a representative speech batch
Use your exact checkpoint, runtime, precision, GPU count, voices and output settings. Time full generation, including any model loading you expect to repeat, and divide elapsed minutes by the generated audio minutes. Record peak memory across a complete run. Leave missing measurements blank to see what still needs testing.
For Drama 3, the community W4A16 example reports 57.9 GB peak GPU memory on an A100 80GB. Its 22.1 GB download is not its runtime memory requirement. The ComfyUI offloading experiment simulated an 8GB limit on an RTX 3090 with 188GB of system RAM. Neither supplies a transferable batch speed preset. See the model page for the exact configurations and sources.
A memory match here means the entered memory peaks fit the available hardware. Throughput, first-audio delay and concurrent listeners are separate measurements; this batch estimate does not size a live voice service.
Existing GPU or rented capacity?
An existing rig defaults to generating and idle electricity. Measure whole-system power at the wall, especially when weights are offloaded to RAM. You can add the share of equipment depreciation allocated to this workload, plus storage or operations costs.
Cloud cost is generating hours plus the reserved idle hours you enter, multiplied by GPU count and the price per GPU-hour. Include time billed while the rig waits between jobs. For a continuously rented rig, generating and idle hours together must cover the full rental period.
The API comparison below uses its own billable character count. Compare the same scripts and takes in each budget; audio hours are not automatically converted to characters. Local electricity-only cost and a managed API invoice cover different services, so add your own operating costs before choosing.
