⚡
Agent Configuration
Inworld Realtime voice agent - mirrors the Sparkssurge agent config page
Inworld Realtime
TTS-2
Twilio bridge
General
Voice
Turn Detection
Backchanneling Listening
Responsiveness Fillers
LLM
Speech
Memory
Tools
Expression
Phone Live
Agent Name
Call Direction
📞 Inbound Agent
Customer calls in. Qualifies inquiries, answers pricing, books studio shoots.
📤 Outbound Followup
Speed-to-lead follow-up, quotation checks, reschedule calls.
Transport
(identical session config for both)
🌐 WebRTC (recommended)
Direct browser peer connection to Inworld. Platform echo cancellation and interruption, lowest latency. Audio flows as RTP tracks.
🔌 WebSocket (proxy, manual audio)
Local Node proxy with manual PCM capture and playback. Kept for A/B debugging and diagnostics.
WebRTC uses one REST call for the session plus an "oai-events" data channel; WebSocket streams PCM through the local proxy.
Initial Greeting Message Starter
(Spoken first upon connect)
Sent as an initial turn trigger so Tracy speaks first naturally without waiting.
Master System Prompt
Reset to Murielle-Approved Prompt
Governs Tracy's personality, objections handling, pricing breakdown, and booking rules.
Inworld Character Voice
282 Voices
Default is Sarah (Warm & Professional). All Inworld character voices are available with 0 extra markup.
Speaking Speed
1.15x
1.0x (Standard)
1.10x (Natural)
1.15x (Recommended)
1.20x (Fast)
Changes audio output speed in real time. Can also be hot-swapped mid-call.
TTS Model & Delivery Settings
TTS Model
✨ inworld-tts-2
Flagship. Steerable emotion tags, non-verbal cues, 200+ languages.
⚡ inworld-tts-2-flash
Ultra-low ~20ms TTFB. Cheaper, lowest latency, unsteered.
Delivery Mode (TTS-2 Preset)
Conversational Multi-Turn Context
Synthesizer conditions agent audio on caller's prior turn audio for natural tonality matching.
TTS Segmentation & Timestamps
LLM Token Segmenter Strategy
Controls how tokens from the LLM are broken into TTS synthesis jobs.
Steering Tag Handling
Timestamp Alignment (Lip-Sync / Captions)
Type:
Transport:
Turn Detection Engine
🧠 Semantic VAD (AI Predictor)
Analyzes phrase completeness and grammar so it does not interrupt mid-sentence.
🔊 Server VAD (Acoustic Silero)
Pure acoustic silence and voice activity thresholding.
VAD Eagerness / Sensitivity
Barge-In & Interruption Behavior
Caller Barge-In (Interrupt Response)
Cut off Tracy immediately if the caller speaks while she is talking.
Auto Create Response on User Turn
Automatically trigger LLM response when user finishes speaking.
Fine-Grained Threshold Overrides (providerData.stt)
End-of-Turn Confidence Threshold
0.60
STT confidence cutoff to end user turn. 0.60 delivers snappy, ultra-low latency completion.
Acoustic VAD Threshold
0.50
Speech/silence cutoff. Higher values reject background noise.
Min End-of-Turn Silence (ms)
Minimum silence ms before checking turn end (160ms for instant responses).
Max Turn Silence Fallback (ms)
Hard silence ms cutoff that ends turn regardless (1200ms prevents dead air).
💬 Active Listening Backchanneling: Emits short, low-latency interjections ("uh-huh", "right", "I see") while the caller is still speaking. Gives the agent the authentic feel of an attentive human listener without interrupting the user.
📖 Inworld Handbook Natural Backchanneling Rules:
To keep listening interjections completely natural and stop the agent from over-interjecting:
• Require Smart-Turn Pause: Enabled (only nods during vocal pauses, never speaks over the caller mid-word).
• Max Per Turn: 1 (prevents annoying machine-gun nodding).
• Min Speech Before Fire: 2200ms (suppresses interjections on short 1-second utterances).
• Min Gap Between Fires: 6000ms (ensures generous breathing room between interjections).
• Volume Gain: 0.40x (subtle background volume so it never overpowers the customer).
To keep listening interjections completely natural and stop the agent from over-interjecting:
• Require Smart-Turn Pause: Enabled (only nods during vocal pauses, never speaks over the caller mid-word).
• Max Per Turn: 1 (prevents annoying machine-gun nodding).
• Min Speech Before Fire: 2200ms (suppresses interjections on short 1-second utterances).
• Min Gap Between Fires: 6000ms (ensures generous breathing room between interjections).
• Volume Gain: 0.40x (subtle background volume so it never overpowers the customer).
Enable Backchannel Responses
Emit listening interjections mid-utterance while user produces speech.
Decider Kind
🧠 LLM Decider (Default)
Small fast model evaluates partial transcript to pick natural interjections.
🎲 Rule Decider
Picks random phrases from bank based on fire probability without LLM latency.
Volume Gain Multiplier
0.40x
Handbook: Backchannels should be subtle and soft (0.35x-0.45x) so they never compete with the caller.
Evaluation Interval (ms)
How often manager checks for interjection opportunity.
Min Speech Before Fire (ms)
Handbook suggests 2200ms for production; 1500ms here so short test sentences can trigger it.
Min Gap Between Fires (ms)
Handbook: 6000-8000ms ensures interjections are infrequent and natural.
Max Per Turn
Handbook: Max 1 per turn prevents repetitive machine-gun nodding.
Require Smart-Turn Pause (Handbook Recommended)
ONLY fire after an acoustic pause signal. Eliminates mid-sentence vocal collisions completely.
Allowed Phrase Bank (allowed_phrases)
Aussie Style
Professional Studio
Casual & Friendly
⏳ Intermediate Fillers (Dead-Air Masking): Bridges the latency gap when the main LLM is slow to produce its first delta or executing complex tools. A small filler model speaks a brief phrase ("let me think", "one moment") then transparently hands off to the main response.
Enable Responsiveness Fillers
Speak short intermediate fillers if the main LLM takes too long.
Initial Wait Timeout
1200ms
How long to wait for the main LLM's first delta before triggering the filler.
Hard Deadline (ms)
Caps the filler model's streaming time.
Min Filler Gap (ms)
Spacing between any two fillers.
Allow Filler on First Reply
Permit fillers on the opening greeting (usually false so greeting plays clean).
Pause Connector Text (pause_text)
TTS-only hint injected between the filler and the main answer.
Language Model (At Cost Router)
31 Models
All router models billed at raw provider cost with 0 markup. Anthropic Claude (Opus 5, Sonnet 5, Haiku 4.5, Fable 5.1), OpenAI (Astra, Sol, 4.1), and Gemini fully verified.
Temperature
0.70
Lower is more factual & strict; higher is more conversational and spontaneous.
Chain-of-Thought Reasoning
Reasoning Effort (Supported Models)
Controls thinking tokens generated before speech synthesis begins.
Reasoning Max Tokens
Cap on reasoning tokens.
Max Output Tokens
Caps speech turn length.
Exclude Reasoning from Spoken Audio
Generate reasoning tokens for accuracy, but omit them from audible synthesis.
Prompt Caching (providerData.caching)
Enable Prompt Caching
Attaches cache breakpoints to system instructions and tools to slash input token latency and cost.
Cache TTL
Top P Sampling
STT Speech Recognition Model
Streaming STT model for caller audio. AssemblyAI u3-rt-pro is what Inworld's own playground uses; Inworld STT v1 is the native alternative.
Note: the providerData.stt silence thresholds below and the transcription settings (language, custom vocabulary prompts) are documented for Inworld STT and may be inert while AssemblyAI is selected.
Speech Recognition Language
Acoustic Voice Profiling (providerData.stt.voice_profile)
Enable Real-Time Voice Profile Extraction
Analyzes caller audio and attaches estimated Age, Emotion, Pitch, Vocal Style, and Accent to every transcript event.
Custom Vocabulary Boosting (prompts[])
Guides STT to transcribe custom proper nouns, company names, and local suburbs accurately.
🧠 Automatic Conversation Memory: Periodic summarization and durable fact extraction that keeps long sessions inside the context window by rolling up older exchanges into high-salience bullets.
Enable Automatic In-Session Memory
Automatically extract facts and rolling summaries every N completed turns.
Turn Interval
Generate memory every N turns.
Max Facts Retained
Max facts stored in state.facts.
Max Rolling Summary Length
Character ceiling for summary.
Max Transcript Items Kept
Items kept after trimming.
Knowledge Base Snippets & Context
Context Snippets
VIVID Pricing
ASB Solar Specs
Clear
Injected directly into system context so the model can quote verified facts accurately.
Native Toolset Simulator
Inworld uses OpenAI wire-compatible tool schemas. When Tracy calls a function, the local playground executes the mock endpoint and returns the result in real time.
Auto Tool Response (providerData.auto_tool_response)
Inworld automatically continues with follow-up assistant speech after tool output is sent.
📅 check_availability
Checks morning studio session openings for natural window lighting.
✅ book_appointment
Books a confirmed corporate branding or portrait session at Bossley Park.
↩️ transfer_call
Warm transfer to studio director Murielle Sassine or senior photographer.
⛔ end_call
Conclude the call smoothly after booking inquiry or pricing is answered.
⚡ Inworld TTS-2 Persistence Rule
Emotion tags like
[warm] or [calm] persist across every subsequent sentence until explicitly changed or cleared with [reset]. Non-verbal tags like [laugh] or [sigh] trigger once at the exact position in the sentence.
Persistent Emotion Tags (Click to copy)
😊 [warm]
🌊 [calm]
🤔 [curious]
⏳ [patient]
✨ [happy]
💪 [confident]
🤫 [whispering]
🎉 [excited]
🧐 [serious]
💼 [professional]
🔄 [reset]
Non-Verbal Audio Cues (One-off)
😂 [laugh]
😮💨 [sigh]
😲 [gasp]
🤧 [cough]
🔊 [clear_throat]
🐽 [snort]
🥱 [yawn]
📞 Twilio phone bridge: Twilio Media Streams pipe caller audio straight through to Inworld, which accepts G.711 mu-law (audio/pcmu) at 8 kHz natively - there is no transcoding in the path. Phone calls always use the saved default config, so press Save as default after changing the controls below.
Twilio number
checking...
The number callers dial. Inbound calls reach this playground once the number's voice webhook points at the URL below.
Voice webhook URL (set this on the Twilio number)
HTTP POST. The TwiML returned there opens the Media Stream to /media-stream
Test call (outbound)
Outbound test calls pass the voice URL explicitly, so the number's stored webhook does not need changing for them to work.
Recent phone calls (0 of last 20)
No phone calls yet. Place a test call or dial the Twilio number.
Test call
Live audition
🎙️
Tracy - VIVID Photography & Imaging
Ready to audition Tracy on Inworld TTS-2
Press "Start Live Call" to open a bidirectional Inworld Realtime session.
Speak into your microphone to hear Tracy reply with low-latency TTS-2 and active backchannels.
Speak into your microphone to hear Tracy reply with low-latency TTS-2 and active backchannels.