Self-host ONNX TTS vs Audra API: ops cost honesty — Audra

Piper and other ONNX-exported models have open weights: the model itself costs nothing. The comparison that matters is total cost of ownership at your volume — GPU time, engineer hours, QA, and uptime. Here is an honest model at a mid-size batch workload.

What self-hosting actually costs

• GPU hours: 2M characters is roughly 44 hours of finished audio; a T4-class GPU renders near-realtime, so budget ~45 GPU-hours per month before retries.
• A machine that is up when your cron fires — spot instances go away at the worst moment, so most teams pay for reserved or on-demand uptime.
• Engineering time: model selection, chunking, loudness normalization you now own, failed-render handling you now own, upgrades when weights move.

Worked TCO: 2M characters a month

A small reserved GPU instance runs $60–90/month; add redundant capacity or burst handling and the realistic floor for 1M monthly characters is ~$110/month before anyone's salary. Audra's checkout price for the same volume is $25 — flat, regardless of spend, because tiers are still planning-only.

When self-host genuinely wins

• You already run idle GPU capacity for other batch jobs — marginal cost of a render approaches zero.
• Compliance requires the model inside your VPC and no text leaves the boundary.
• You are past 20M chars/month and can staff the pipeline as an internal platform.

What you trade away either way

Hosted, you get QA-gated refunds on failed renders, −16 LUFS mastering in the export, a status page, and engine upgrades without redeploying. Self-hosted, you own all of it — and that is the honest trade. Architecture notes are at /blog/why-audra-essential; pricing at /pricing.

Next steps

• Start here: /docs/engine
• Then: /blog/why-audra-essential
• Reference: /pricing
• More Audra guides: /blog