Conversational AI Advanced Settings Guide

NSPECT Conversational AI allows your Spark to have natural conversations with callers using voice AI. In most cases, the default settings work extremely well and should not need to be changed.

This guide explains what each setting does, when you might want to adjust it, and what impact those changes will have on the caller experience.

Our recommendation: Start with the default settings and only make changes if you notice a specific issue.

For most phone-based Sparks:

These settings provide the best balance of reliability, call quality, and natural conversation for most users.

What it does

Voice determines how the AI sounds when speaking to callers.

When to change it

You may want to experiment with different voices to better match your company brand or audience.

Example

A utility company may prefer a professional voice, while a customer service Spark may benefit from a warmer, friendlier voice.

Recommendation

Leave blank to use the system default voice (Alloy).

Recommended Value: 0.8

What it does

Temperature controls how creative and varied the AI’s responses are.

Think of it as a personality slider.

• Lower values = More consistent and predictable
• Higher values = More conversational and natural

When to increase it

Increase the temperature if conversations feel robotic or repetitive.

When to decrease it

Decrease the temperature if the AI starts giving overly creative answers or becomes inconsistent.

Examples

Temperature 0.3

• Very consistent
• More scripted
• Good for compliance-driven workflows

Temperature 0.8

• Natural conversation
• Friendly responses
• Recommended for most Sparks

Temperature 1.0

• Most creative
• Highest variation
• Can occasionally be less predictable

Recommendation

Use 0.8 for most conversational experiences.

Recommended Value: server_vad

What it does

Detection Mode determines how the AI decides when a caller has finished speaking.

You can think of it as the AI’s “listening style.”

server_vad

Best For

• Phone calls
• Noisy environments
• Speakerphones
• General use

How it works

The AI listens to audio volume levels and pauses to determine when someone has finished talking.

Benefits

• Extremely reliable
• Works well with poor phone connections
• Handles background noise better

Recommendation

This is the best option for most Sparks.

Best For

• High-quality audio
• Advanced conversational experiences

How it works

The AI uses language understanding to determine when someone has completed a thought.

Benefits

• More natural conversations
• Fewer interruptions
• Better flow when callers pause briefly

Considerations

May be less reliable in noisy environments or poor phone connections.

Recommended Value: Off

What it does

Controls whether callers can interrupt the AI while it is speaking.

Off (Recommended)

The AI finishes its sentence before listening again.

Benefits

• Prevents accidental interruptions
• Reduces issues caused by speakerphone echo
• More stable phone call experience

Best For

Nearly all phone-based Sparks.

On

Callers can begin speaking while the AI is talking.

Benefits

• Faster conversations
• More human-like interaction

Risks

• Background noise may accidentally interrupt the AI
• Speakerphone echo can cause unwanted interruptions

Recommendation

Leave Off unless you have a specific reason to allow interruptions.

Only Available When Using semantic_vad

What it does

Controls how quickly the AI decides that a caller has finished speaking.

Think of it as the AI’s patience level.

Low

The AI waits longer before responding.

Benefits

• Fewer interruptions
• Better for phone calls

Best For

Most users.

Medium

Balanced waiting time.

Best For

General conversations.

High

The AI responds very quickly.

Benefits

• Faster conversations

Risks

• May interrupt callers who pause while thinking

Recommendation

Use Low for phone-based Sparks.

Recommended Value: 0.85

What it does

Controls how sensitive the AI is to sound.

Higher Values

Examples: 0.85–1.0

Benefits

• Ignores more background noise
• Fewer false activations
• Better for noisy environments

Tradeoff

May miss very quiet speakers.

Lower Values

Examples: 0.50–0.70

Benefits

• Picks up quieter voices

Tradeoff

More likely to react to background sounds.

Recommendation

Use 0.85 unless callers are speaking unusually softly.

Recommended Value: 900 ms

What it does

Controls how long the AI waits after a caller stops speaking before responding.

Longer Values

Examples: 1200–1500 ms

Benefits

• Gives callers more time to think
• Reduces interruptions

Best For

Older callers, complex questions, or slower conversations.

Shorter Values

Examples: 500–700 ms

Benefits

• Faster conversations

Risks

• May feel rushed
• Can interrupt callers who pause briefly

Recommendation

900 ms works well for most phone conversations.

Recommended Value: 500 ms

What it does

Captures a small amount of audio before speech is detected.

This helps prevent the beginning of a sentence from being cut off.

Example

Without enough padding:

“…need help with my account.”

Instead of:

“I need help with my account.”

When to increase it

If callers report that the AI is missing the first word or two of what they say.

Recommendation

Leave at 500 ms unless directed by NSPECT support.

The AI is interrupting callers

Try:

• Increase Silence Duration
• Use server_vad
• Set Eagerness to Low
• Turn Allow Interruptions Off

The AI takes too long to respond

Try:

• Reduce Silence Duration
• Increase Eagerness (semantic_vad only)

Background noise is causing problems

Try:

• Increase Threshold
• Use server_vad
• Turn Allow Interruptions Off

The AI sounds robotic

Try:

• Increase Temperature slightly
• Test a different Voice

For 95% of Sparks, these settings provide the best caller experience:

• Voice: Alloy (Default)
• Temperature: 0.8
• Detection Mode: server_vad
• Allow Interruptions: Off
• Threshold: 0.85
• Silence Duration: 900 ms
• Prefix Padding: 500 ms

If you’re unsure, start with the defaults and make small adjustments one setting at a time so you can clearly hear the impact of each change.