preferred SBL1
checking - model
- —
- device
- —
- compute
- —
- service
- —
- latency
- —
- cpu
- —
- ram
- —
- disk
- —
- load
- —
- uptime
- —
speech-to-text
voice
fallback SBL4
checking - model
- —
- device
- —
- compute
- —
- service
- —
- latency
- —
- cpu
- —
- ram
- —
- disk
- —
- load
- —
- uptime
- —
input ready
Release to transcribe automatically
result transcript
- backend
- sbl1
- model
- —
- elapsed
- —
request waiting
- duration
- —
- router
- —
- attempts
- —
about
local voice inference path
The browser records a short microphone clip with the built-in MediaRecorder API. When the button is released, the audio bytes are sent through the authenticated too.foo audio bridge, not to a third-party speech API.
The bridge forwards the clip to the SBL speech-to-text router. The router prefers SBL1 for the CUDA large-v3 path and can fall back to SBL4 for the Vulkan path. The cards above show only operational metadata that is safe for a demo: health, model label, device class, compute mode, and latency.
Secrets stay server-side. This page does not contain Cloudflare tokens, SSH keys, model paths, service credentials, or internal filesystem details. The public app only sends the recorded audio clip and receives transcript text plus sanitized backend metadata.