🏆 1st Place ElevenLabs Hackathon – $20,000🚀 EBRD Star Venture Program🥈 2nd Place Sevan Startup Summit🚀 Google Cloud $25K Grant
Kallina AI
RO
Kallina Voice AI

Echo Cancellation

Eliminare echo pentru conversații clare și naturale.

Clear Conversations, No Echo

Echo-ul degradează conversația și poate confuza STT-ul. AEC (Acoustic Echo Cancellation) elimină feedback-ul audio.

Types of Echo

Acoustic Echo

Sound from speaker picked up by microphone in same room.

🔊🎤

Speaker output → Room acoustics → Microphone input

Common with speakerphones, laptops, smart speakers

Line (Hybrid) Echo

Signal reflection at 2-wire/4-wire conversion points in PSTN.

4-wireHybrid2-wire

Impedance mismatch causes signal reflection

Common in PSTN gateways, international calls

How AEC Works

1

Reference Signal Capture

AEC captures the audio being sent to the speaker (far-end signal).

2

Estimate Echo Path

Adaptive filter learns how far-end signal appears in microphone input (room acoustics, delays, frequency response).

3

Generate Anti-Echo

Create inverted estimate of echo component.

4

Subtract from Microphone

Remove echo while preserving near-end speaker's voice.

AEC Signal Flow

Far-end audio (from caller)
       │
       ▼
   ┌───────┐
   │Speaker│ ──────────────────────────┐
   └───────┘                           │ (acoustic path)
                                       ▼
   ┌───────┐     ┌─────────┐     ┌──────────┐
   │  Mic  │ ──► │  AEC    │ ──► │ Clean    │ ──► To STT/Recording
   └───────┘     │ Filter  │     │ Output   │
       ▲         └────┬────┘     └──────────┘
       │              │
       │         Reference signal (speaker output)
       │
  Near-end speaker (local user)

AEC Parameters

Tail Length

How far back to look for echo

128ms
Suppression Level

Aggressiveness of echo removal

High
Convergence Rate

How fast filter adapts

0.5
ERLE Target

Echo Return Loss Enhancement

>30dB

Echo Metrics

42dB
ERLE
Echo suppression
0.3%
Residual Echo
Remaining after AEC
78ms
Echo Delay
Path delay detected
98%
Convergence
Filter stability

Double-Talk Detection

When both parties speak simultaneously, AEC must not remove the near-end voice:

Far-End Only

Only caller speaking. AEC adapts and subtracts echo.

Normal operation

Near-End Only

Only local user speaking. AEC passes through.

Pass-through mode

Double-Talk

Both speaking. AEC freezes adaptation to protect near-end.

Freeze & protect

Voice AI Considerations

Why Echo Matters for Voice AI

  • • STT hears AI's own speech → transcription errors
  • • Barge-in detection fails
  • • User frustration from hearing themselves
  • • Call quality scores drop

Best Practices

  • ✓ Apply AEC before STT input
  • ✓ Use reference signal from TTS output
  • ✓ Monitor ERLE in production
  • ✓ Test with various device types

WebRTC AEC

// Enable WebRTC echo cancellation
const constraints = {
  audio: {
    echoCancellation: true,
    echoCancellationType: 'system', // or 'browser'
    autoGainControl: true,
    noiseSuppression: true,
  }
};

const stream = await navigator.mediaDevices
  .getUserMedia(constraints);

// Check if AEC is actually enabled
const track = stream.getAudioTracks()[0];
const settings = track.getSettings();
console.log('Echo cancellation:', settings.echoCancellation);

Crystal Clear Voice

Echo cancellation pentru conversații fără distorsiuni.

Vezi Demo →
Începe Astăzi

Transformă Comunicarea cu Clienții

Agenți vocali AI care răspund 24/7 în română și rusă. Implementare în 2 săptămâni, fără infrastructură specială.

Setup în 24 oreSuport dedicatGDPR compliant

Rămâi la curent

Obține cele mai recente știri despre tehnologia de apelare AI și actualizările platformei

Made with ♡ by Kallina AI Team — 2025