Saved on the server in dataDir/runtime.json (file mode 600, ignored by git) and kept across reloads and restarts. The browser never stores the secret.
Raw session (advanced)
Audio source:
PCM16 WAV, mono or stereo, any rate (left channel, FIR resampled to the selected media), or raw .ulaw/.raw μ-law 8 kHz. Up to 20 MB.
0.0sThe microphone only works on localhost or HTTPS. Echo cancellation, noise suppression and AGC are off to keep the raw voice.
Selected audio: (none)
inputVariables:
Media offers:L16 is sent as little-endian PCM16.
Scenarios:
Call journey
One row per attempt, grouped by conversation. ✓ done · ● now · ✗ failed · – skipped · ○ not yet.
Conversation
#
Inputs
Connect
open → opened
Caller audio
Server prompt
Result
close → closed
Now
no call yet
Sessions (the last 100 finished sessions are kept)
sid
conversationId
state
t (s)
audio (s)
RTT
disconnect
outputVariables
Actions
no session yet
Raw JSON
Raw JSON → session:version/id/seq/serverseq/position are filled in when missing. A seq you write is sent as is and does not move the counter.
Manual test
Action:
liveRuns the whole flow (Audio Connector, re-prompt on too_short/low_quality/multi_speaker/TIMEOUT, limit of 5), checks the BRD output and saves every attempt to the customer history.
Message exchange
Pick a call here or click a row in Call journey. Each dot is one log line, the arrow shows who sent it. Click a line to see its JSON.
Pick or record audio in "Raw session (advanced)" of the Test tab, then assign it. voice_short is the first 2 s of voice_good; silence and noise need no file. voice_other is also the agent on the internal channel (P2, P3).
Run settings
Connector ID:
Input/output contract (BRD 6.1 and 7, JSON override)
Customers & report (synthetic customer registry)
Synthetic data only: never real customers or real voices.