Live metrics from the server — use this to check if the second concurrent call is being throttled by DashScope. If LLM connect or First token jumps a lot when Active Sessions > 1, the API is rate-limiting.
Active Sessions
0
LLM Connect
--
LLM First Token
--
First Audio
--
TTS Wait
--
LLM Total
--
TTS Drain
--
Total
--
Turn history (last 10) — look for rising times:
#
Sess
LLM Conn
1st Tok
1st Audio
LLM Total
Total
No turns yet — say something to start collecting data.
Agent A — Primary
▲
Spoken via TTS on first contact — skips LLM, so it plays instantly. Leave blank for AI-generated greeting.
Only used when the Language Agent is enabled. Leave blank to reuse the First Greeting on return.
Language Agent
▲
Prepends a language-selection step before the main agent. Uses the Voice Mapping keys as the language options and the English voice for its own speech. Once chosen, the whole session is locked to that language.
Spoken at call start (and again if the caller asks to change language). Plays instantly via TTS.
Leave blank to use the default prompt auto-built from Voice Mapping keys.
Voice Mapping
▲
ASR auto-detects the spoken language. Map each language to a TTS voice.
Allowed Languages
▲
Select which languages the system will accept. Callers speaking an unchecked language
will hear a polite rejection. If none are checked, all Voice Mapping languages are allowed.
Voice Cloning
▲
Clone a voice from 10-20 seconds of clear audio. Cloned voices appear in Voice Mapping dropdowns.
Click to upload or drag & drop audio (.wav, .mp3, .m4a)
WAV/MP3/M4A, 10-20 sec of clear speech, no background noise, mono ≥24kHz
TTS Settings
▲
1.0
0.5 (slow) — 2.0 (fast), default 1.0
1.0
0.5 (low) — 2.0 (high), default 1.0
50
0 (silent) — 100 (max), default 50
Test TTS with current settings
Agent Data / Knowledge Base
▲
Click to upload or drag & drop .txt files
Text content is appended to the system prompt as reference data
Call Variables
▲
Use [VARNAME] placeholders in any agent prompt or tool definition. They are replaced with live call values at runtime.
Placeholder
Last Value
No call data yet
Map config fields from Asterisk to placeholders. Enter a placeholder name to activate.
Config Field
Last Value
Placeholder
Fallback
Audio Settings
▲
Plus is more accurate for complex prompts and tool calling. Flash is faster and cheaper.
Voices for qwen3.5-omni. All support Chinese + English + more.
Adjust AI voice volume (0-200%). Does not affect recording.
Injected into Omni system prompt. Controls pacing, tone, and delivery style. Leave empty to skip.
Voice energy threshold for barge-in. Lower = more sensitive.
When the caller interrupts mid-sentence, nudge the AI to continue the
flow instead of restarting. Turn off to revert to plain interruption handling.
Denoise the caller audio before it reaches the model (helps in noisy
environments; DashScope VAD unchanged). Off = exact current behavior. The call recording
reflects this setting. Requires the GTCRN model on the server; falls back to Off if absent.
Speech-to-text engine. Gemini requires GOOGLE_API_KEY env var on server.
Select a known model or type a custom model ID
Select none for auto-detect. First selected language is used for speechConfig bias; all are added to system instruction.
0.20
Higher = harder to interrupt / ignores quiet speech & noise. Lower = more sensitive. DashScope range -1.0 to 1.0 (API default 0.5). In Omni mode this is the server-side turn-detection threshold.
How long the caller must pause before the AI responds (200 - 6000; DashScope suggests 500-600 for short answers). Higher = fewer premature cut-ins.
semantic_vad filters backchannels ("uh-huh") and noise; silence duration becomes the patience ceiling. Supported on qwen3.5-omni-* and qwen3.8-omni-flash. Prefix padding and idle timeout apply to server_vad only.
Audio kept before detected speech so the first syllable is not clipped. Blank = DashScope default (300).
qwen3.5-omni-* only: if the caller stays silent this long after playback ends, the model speaks to re-engage them (5000 - 30000). Blank = off. Not sent for other models.
These turn-detection values are saved with the profile and are final for phone calls: the bridge's connect-time VAD values (it sends silence 500 ms) are ignored where the profile sets a value.
Barge-in (caller interrupts AI playback) is always on. This is the minimum speech length that triggers it (filters noise).
Prompt the caller after silence, then end call if no response
Seconds of silence before each prompt (3 prompts total before hangup)
LLM Settings
▲
0.70
1.00
Functions / Tools
▲
Built-in tools (click to toggle):
Transfer Agents
▲
Define up to 10 departments/agents that callers can be transferred to.
Silent pause between farewell and greeting when transferring between agents
ACRS Dispatch
▲
When enabled, conversations are automatically classified and dispatched to ACRS after each session.
Departments & Categories
Types
Subcategories
Ratings
Callback Defaults
Pre-select ACRS defaults for callback requests from transfer agents. Category and subcategory are auto-classified from the conversation. Leave Department empty to skip ACRS dispatch on callbacks.
Custom Payload Fields
Extra key:value pairs added to every dispatch payload.
Conversation Validator
▲
When enabled, the full conversation history is POSTed to the configured URL when a session ends.
Emotion Detection
▲
Parallel emotion analysis via emotion2vec. Requires a separate GPU service running emotion-engine/emotion_server.py.
Conversation History
▲
No conversations yet
Press Start to begin
Angry
Disgust
Fear
Happy
Neutral
Sad
Surprise
MIC
Advanced Settings
AI Tool Builder
Describe Your Tool
Describe what API tool you want in plain English. Include the endpoint URL,
parameters, and purpose.
Model used for generating tool definitions and handler code.
Tool Definition
Execution Mode
API Endpoint
Custom Handler Code
Python handler function that runs on the server when this tool is called.
Uses @register_handler for automatic registration.
Parameters
Parameters the agent will collect from the caller before calling the API.