Call Quality & Latency

This guide covers diagnosing and resolving audio quality, latency, and conversational flow issues.

​

Diagnosing Latency

Response latency is the time between when a caller finishes speaking and when the agent starts replying. The agent editor toolbar shows estimated latency for your configuration.

​

Latency Breakdown

Total latency = Transcription + AI Model + Voice Synthesis + Network

​

When Latency Is a Problem

• Under 1 second: Excellent — feels like a natural conversation
• 1-2 seconds: Acceptable for most use cases
• 2-3 seconds: Noticeable — consider optimizing or adding thinking sounds
• Over 3 seconds: Poor experience — take action now

​

Quick Fixes for High Latency

1. Switch to a faster model — Go to General → Thinking and try Balanced or Fast presets
2. Enable thinking sounds — General → Sounds → Thinking Sounds (Expert Mode) fills processing time with keyboard audio
3. Enable smart filler — General → Sounds → Smart Filler (Expert Mode) generates contextual filler phrases
4. Use a faster voice provider — Check Supported Providers for latency benchmarks
5. Simplify your prompt — Shorter prompts process faster
6. Adjust VAD turn detection — VAD settings can have a significant impact on perceived response timing

---

​

Turn-Taking Issues

​

Agent Talks Over the Caller

Symptoms: Agent starts speaking while the caller is still talking, or responds too quickly after brief pauses. Fix:

1. Open VAD Turn Detection settings
2. Switch to a more Patient response timing preset
3. In Expert Mode, increase Silence before responding (e.g., from 300ms to 500ms)
4. Enable AI Turn Detection (Expert Mode) for smarter end-of-turn detection

​

Agent Waits Too Long to Respond

Symptoms: Awkward silences after the caller finishes speaking. Fix:

1. Switch to a more Responsive timing preset
2. In Expert Mode, reduce Silence before responding
3. Check if AI Turn Detection is causing delays — try toggling it off
4. Switch to a faster AI model

​

Caller Can’t Interrupt the Agent

Symptoms: Caller speaks but the agent continues its response without stopping. Fix:

1. In Expert Mode, enable Allow Interruptions
2. Reduce Speech duration to trigger interrupt (how long the caller must speak to interrupt)
3. Reduce Minimum words to interrupt threshold

---

​

Audio Quality Issues

​

Robotic or Unnatural Voice

1. Try a different voice — some voices sound better for conversational use
2. Adjust voice settings (Expert Mode) — tweak stability, similarity, and style parameters
3. Lower the Response Style (temperature) setting — higher values can produce less consistent speech patterns

​

Echo or Feedback

• This typically occurs on web calls. Check that the caller’s browser has echo cancellation enabled
• Reduce ambient sound volume if using background audio
• Test with headphones to isolate the issue

​

Muffled or Unclear Speech

1. Check your transcriber selection — Deepgram Nova-3 has the best general accuracy
2. Add custom pronunciations for frequently misheard terms
3. In Expert Mode, add keywords to boost recognition of specific terms

​

Phone vs Web Quality Differences

Phone calls go through additional compression and network hops, which can affect quality:

• Always test with real phone calls before launching — web simulator results may differ
• Phone networks add 50-200ms of latency that web calls do not have
• Codec compression can affect voice quality — some voice providers handle this better

---

​

Silence Handling

​

Agent Does Not Respond When Caller Goes Silent

Configure Inactivity Timeout:

1. Enable Silence Reminders to prompt the caller after a period of silence
2. Set Max Call Duration to end calls that run too long
3. In Expert Mode, configure reminder timing (delay and max count)

​

Agent Hangs Up Too Quickly

1. Increase Max Call Duration (default may be too short for your use case)
2. Check that Allow AI to Hang Up is not ending calls prematurely
3. Review your prompt — remove language that tells the agent to end calls aggressively

---

​

Testing Methodology

1. Start with web calls — fastest iteration cycle, no phone network variables
2. Move to phone calls — test real network conditions, AMD behavior, and voice quality
3. Test from different environments — quiet office, noisy space, mobile, landline
4. Compare models — try the same conversation with different AI models to find the best speed/quality tradeoff
5. Review conversation timelines — check tool execution times and knowledge retrieval latency

​

Next Steps