Assistant Best Practices

Use this simplified guide to choose the right settings, models, and configurations for peak performance.

1. Pick the Right Mode

Dualplex (Best Default)

Fast + high-quality ElevenLabs voice. Works great for Sales, support, appointment setters, and branded voice.

Speech-to-Speech

Fastest + most natural conversational flow. Works great for short, reactive, "chatty" conversations.

Pipeline

Slowest but strongest for logic. Works great for complex reasoning, long explanations, and cloned voice accuracy.

2. Pick a Transcriber (Pipeline Only)

This affects how accurately the assistant hears the caller.

Azure: Highest accuracy, slower. Best for noisy callers or accents.
Gladia: Good accuracy, fast. Best for general use.
Deepgram: Good accuracy, fast. Best for English-heavy users.

Tip: Test all 3 with a short call—keep whichever gives best results.

3. Choose the Right AI Model

If SPEED matters: Use GPT-5 Realtime or Gemini Flash 2.0 / 2.5.

If REASONING matters: Use GPT-4o or GPT-5 Mini (Pipeline mode).

Note: If using slower models, enable filler audio to hide latency.

4. Noise Cancellation

ON: Better hearing in noisy environments.
OFF: More raw sound, better for quiet callers.

Turn OFF if caller audio seems "clipped" or it keeps mishearing.

5. Conversation Timers

Setting	Suggested	Why
Re-engagement Time	30 sec	Gives caller time to think
Max Silence	60 sec	Prevents accidental long pauses

6. Initial Message Strategies

Pipeline / Dualplex: Reads exactly what you type. Write it like a script.

Speech-to-Speech: Interprets it like a suggestion. Use say exactly: if you want a precise greeting (e.g., "say exactly: Hello, this is Sarah with ABC Company.").

7. Endpointing Sensitivity (Who Talks First)

Controls interrupt timing.

Lower Sensitivity: Assistant jumps in faster. Best for quick, snappy conversations.
Higher Sensitivity: Assistant waits longer. Best for long, detailed caller replies.

8. Debugging

If something feels off:

Open Call History
Click the call
Check transcript for misheard words, interruptions, or prompt errors

This is the fastest way to locate a configuration issue.