A voice app may combine transcription, a text model, and speech generation. A realtime speech model may have another billing structure entirely. The length of a conversation is useful product information, but it does not replace the service’s billable units.
Separate the stages
Use an audio-minutes component for minute-priced transcription, a text-token component for responses, and a character component for character-priced speech. For token-priced audio, select a custom audio-token profile and enter the actual input and output rates. The calculator does not guess how many audio tokens fit in a minute.
Do not treat a provider’s estimated minute-equivalent for a token-priced model as an exact invoice formula. Measure your actual language, speaking speed, output length, and silence handling.
A worked conversation
Consider an illustrative five-minute interaction. Enter transcription at $0.01 per minute: $0.05. Enter a text-model response cost of $0.02 for the total conversation. Enter 2,000 generated characters at $10 per million characters: $0.02. The combined cost is $0.09.
Thirty such conversations cost $2.70 monthly for a normal user. A customer with five times the activity costs $13.50. At 10% heavy users, expected cost is $3.78 per paying customer, before free users and other expenses.
If transcription records all input duration but speech generation charges only generated characters, multiplying both by conversation minutes creates the wrong model. Keep the units visible in your workflow.
Check infrastructure separately
Telephony, streaming transport, recording storage, and other services may add costs beyond the model. Enter predictable per-active-user expenses under other costs or use custom components for usage-dependent charges. No live microphone, audio upload, or provider connection is necessary to plan the economics.
Start a voice scenario and verify the selected billing profile.