The use case is unforgiving by design: ELSA delivers real-time pronunciation feedback to 92M+ learners — a speech-to-feedback pipeline that runs on every utterance, in seconds, across every market ELSA operates in. Every one of those utterances is audio that has to be stored, served, and pulled back to improve the models.
The audio corpus is the asset — and on S3, every read to retrain or evaluate a model is an egress bill. That's the door that opens first:
app.elsaspeak.com's current S3/CloudFront path, each of those reads is a per-GB egress charge that scales with how often you iterate. R2 charges zero egress — so the more you retrain, the more it saves. At ELSA's audio volume, that's a structural cost line, not a rounding error. R2 overview.
Three primitives map directly to how ELSA stores, serves, and improves speech:
R2 — the audio corpus at zero egress: every utterance stored where the Insights team can pull it back for fine-tuning and eval without a per-GB egress bill, and retire the CloudFront/S3 path on app.elsaspeak.com onto the platform you already run.
Workers AI — pronunciation-feedback inference at the edge, milliseconds from every learner in 100+ markets, fed by R2 without backhaul. Your speech models, run close to the speaker.
VoidZero (Vite, Rolldown, Oxc) — the team behind your Astro front-end joined Cloudflare on June 4. The build tool running elsaspeak.com right now is built by people on the same team that fronts your domain.
Which is closer to the current sprint — the audio-corpus economics on R2 (and what egress is costing you on S3 today), edge inference with Workers AI, or a direct line into the Vite roadmap? 20 minutes to find the right starting point.
The detailed primitive-by-primitive mapping — including the speech-feedback request flow on Cloudflare, the audio-corpus and egress economics math, the Astro/Vite/Rolldown roadmap angle, the per-market regional residency story, and the 90-day rollout — runs about 52 KB of dense technical content.
Read the expanded version →