agent-native audio proxy — measurement bench

Pick a source and fan it across preset × quality × codec. Each tile fetches the real /audio/… proxy URL, reads the live X-Transcode-* response headers, and plays the bytes the proxy actually returned — nothing here is computed client-side. Savings are measured against the passthrough baseline. The default source is an audio-Bible chapter (the primary use case); the longer literary readings show how savings scale on full chapters.

Heads up: audio transcoding runs in a container. On a deploy without the container running (or locally without Docker), variants return PASS / passthrough and match the source size — the tile shows the X-Transcode-Reason so you can see why. Slice 1 wires voice + opus end-to-end; other combinations pass through by design until later slices. Sample sources: the default audio-Bible chapters are the Berean Standard Bible (BSB), narrated by Bob Souer and dedicated to the public domain (CC0), hosted by the publisher at openbible.com; the literary readings are public-domain LibriVox recordings.

voice · opus targets low 8 kbps · 8 kHz medium 16 kbps · 16 kHz high 32 kbps · 24 kHz all mono. These are the recipe targets (the q knob, like 20/50/80 for images); each tile shows the actual measured bitrate. Opus always delivers at 48 kHz decode regardless of coded rate. Other preset/codec combinations pass through unchanged in slice 1.

audio measurement bench read the bytes, play the result