TTS API billing: why QA-gated metering matters for pipelines — Audra
Most TTS APIs meter every request — including silent failures, truncated audio, and retry storms. Audra reserves credits when a render starts, then finalizes only after automated QA passes. Failed gates auto-refund to your balance.
Reserve → render → QA → finalize
POST /v2/speech trims input, reserves billable characters, runs inference + mastering, then checks duration and silence gaps. Credits finalize on pass; otherwise the reservation refunds. Check X-QA-Passed on responses during integration testing.
What a failed gate looks like in a pipeline
A failed render is not an error response — the request completes, credits that were reserved at render start are refunded in full, and the response carries X-QA-Passed: false. Silent gaps and truncated (too-short) audio are the two triggers. Read X-QA-Passed like a HTTP status: log it, count it, and alert on failure-rate spikes instead of auditing audio by ear. A chunk that fails twice is usually a text problem (bad normalization, stray control characters) — fix the input, not the retry.
Why pipelines care
- Nightly cron retries do not drain credits on bad chunks
- Course platforms avoid billing students for broken lesson audio
- Podcast workers can fail loud on QA instead of shipping silence
- Idempotency keys prevent double-billing duplicate POSTs
What QA does not cover
QA validates technical output (duration, silence) — not subjective voice fit or pronunciation taste. Use /dashboard/playground and /voices previews before production cutover. Phonetics tooling on Essential helps finance and medical terms.
Refund policy details at /refund. Pricing and commercial unlock at /pricing. More billing FAQs at /faq. Live stats and savings methodology on /proof.