🎤

Voice Agent

Turns 0

Session —

Connecting…

Live Transcript

Live

Training

Prompting

💬

Waiting for a call to start…

Agent Memory ▶

🎤

Idle

Prompt

Voice

Models

Sessions

Instructions

Settings

Conductor

Agent Name → {agent_name} in prompt

Master Prompt

{distilled_instructions} {training_preamble} {agent_name} ← substituted in-place at runtime

🧪 Prompt Engineering

Language Instruction variable: {language}

Training Preamble injected at {training_preamble} in the master prompt

Mode Label variable: {mode}

Instruction Memory Label legacy header — distilled instructions now inject at {distilled_instructions}

Master Prompt variables: {agent_name} {conversation_history} {sentence_facts}

Live

Prompting

No completed sessions yet

Instruction Memory

No training sessions yet — run a session in 🎓 Training mode to build memory

Voice

💡 Tip: Selecting a voice that matches your session language gives better pronunciation quality.

Or enter a custom ElevenLabs Voice ID:

TTS Model

Voice Settings

Stability 0.50

Higher = more consistent, lower = more expressive

Similarity 0.75

How closely the voice matches the original sample

Style 0.00

Style exaggeration — 0 is neutral (recommended for voice agents)

Speed 1.00

Playback speed — 0.7 (slow) to 1.2 (fast)

Speaker Boost

Enhances clarity for English numerals & speech

STT Provider

STT Model DEEPGRAM

Language

BCP-47 tag — e.g. en, en-US, tr, de, fr

Endpointing (min silence) 300ms

Silence (ms) before STT checks for sentence end — u3-pro: fires immediately if punctuation found

Max Turn Silence (force-end) 1500ms

Force end-of-turn after this silence (ms) even without terminal punctuation — must be ≥ Endpointing

Barge-in Sensitivity 350ms

How long user voice must sustain before interrupting agent — 0 = off

OffAggressiveBalancedConservative

Interrupt Delay 300ms

How long agent continues speaking after barge-in fires — 0 = instant cut

InstantNaturalSoft

Saved Numbers

Quick-dial contacts. These appear in the call modal for one-click dialing.

Twilio — Outbound Calls

Phone number the agent calls from (E.164 format). Overrides the TWILIO_PHONE_NUMBER set in .env.

Session Language

Sets the conversation language — affects speech recognition (STT), voice synthesis (TTS), and instructs the agent to respond in that language.

🇬🇧 English

🇩🇪 German

Current: English · Takes effect on next session start.

Conductor Prompt

Schema & Rules

Template variables: {phrase_list}, {tones} — filled at runtime

Model

Backchannel

Backchannel Lock 3.0s

Cooldown between backchannels — higher = less frequent

Opener Lock 2 turns

Turns between openers — higher = less frequent

Enable Situational Backchannels

Situational phrases ("Got it.", "I see.", "Makes sense." …) fire when the model detects a meaningful moment. Disable if they feel off — listening backchannels ("Mhm.", "Uh-huh.") still play.

Cache

EN follows session language

Long Prosody Elongation 4

How drawn-out "long" variants sound (1 = subtle, 4 = very long)

Loading…

Opener Cache

Pre-recorded opener clips — played instantly at speech_final while LLM generates. Each phrase is generated with its category emotion tag via ElevenLabs v3.

Loading…

Conversation