WebRTC + AI Noise Suppression in 2026: Krisp, RNNoise, and the BVC Pricing Shift

Two things changed the noise-suppression calculus in 2026: Krisp BVC moved to paid usage on May 1, and RNNoise stopped getting maintenance updates from Mozilla. Production WebRTC apps now have to make a real decision — pay for Krisp/Telnyx VIVA, ship RNNoise WASM yourself, or rely on WebRTC's built-in NS3 and accept the quality hit.

Why this matters

WebRTC's built-in noise suppressor (NS3, the Speex/WebRTC AEC3 chain) is fine for two-way human conversation in moderate environments. It is not fine for AI voice agents, where ASR and turn-detection are unforgiving — even 50 ms of background TV chatter can trigger a false interruption, derail a tool call, or push your end-pointing model into a hallucination. In 2026, every serious voice-AI vendor (LiveKit, Daily, Telnyx, Vonage, Twilio) ships a deep-learning suppressor on top of WebRTC's classic chain.

The market has bifurcated. Krisp BVC v3 is the gold standard but introduced metered pricing for SDK consumers in May 2026. RNNoise, originally a Mozilla project that married DSP to a small RNN, still works and compiles cleanly to WebAssembly — but it has not been maintained since 2024 and its model is showing its age on modern noise types (mechanical keyboard chatter, baby cries, fan noise from data centers). Newer entrants — Telnyx VIVA, Picovoice Koala, NVIDIA Maxine — split the difference with mid-tier pricing and on-device execution.

Architecture

```mermaid flowchart LR Mic[Browser Mic] --> AWN[AudioWorkletNode] AWN --> WASM[Noise Suppressor WASM
Krisp / RNNoise / Koala] WASM --> AEC[WebRTC AEC3] AEC --> RTP[RTP Encode Opus] RTP --> Gateway[Pion Go gateway 1.23] Gateway --> Agent[Voice Agent Pod] ```

CallSphere implementation

CallSphere runs noise suppression in two places — server-side for inbound PSTN calls hitting the SIP gateway, and client-side for browser WebRTC sessions:

• Real Estate (OneRoof) — Inbound buyer calls land on the Pion Go gateway 1.23. SIP-arriving audio passes through a server-side suppressor before NATS-forwarding to the 6-container agent pod (CRM, MLS, calendar, SMS, audit, transcript). See /industries/real-estate.
• /demo browser path — The marketing site runs RNNoise WASM in an AudioWorkletNode before audio ever leaves the tab. The same WebRTC pipeline supports Krisp SDK as a paid upgrade for affiliates and enterprise tenants. Try it at /demo.
• HIPAA + SOC 2 — The cleaned audio is logged into the audit pipeline (1 of 115+ tables) only for clinical and legal verticals where transcripts are part of the record-of-care.

Across 37 agents, 90+ tools, 6 verticals, and HIPAA + SOC 2, the same noise pipeline ships to every tenant. Pricing remains $149/$499/$1499; 14-day /trial; 22% affiliate at /affiliate.

Build steps with code

```javascript // 1. Load RNNoise WASM into an AudioWorklet const ctx = new AudioContext({ sampleRate: 48000 }); await ctx.audioWorklet.addModule("/rnnoise-processor.js");

// 2. Build the graph: mic -> RNNoise -> destination -> RTCPeerConnection const stream = await navigator.mediaDevices.getUserMedia({ audio: { echoCancellation: true, // keep WebRTC AEC3 noiseSuppression: false, // disable WebRTC NS3 (we use RNNoise) autoGainControl: true, }, }); const source = ctx.createMediaStreamSource(stream); const rnnoise = new AudioWorkletNode(ctx, "rnnoise-processor"); const dest = ctx.createMediaStreamDestination(); source.connect(rnnoise).connect(dest);

// 3. Add the cleaned track to the peer connection const pc = new RTCPeerConnection({ iceServers }); dest.stream.getAudioTracks().forEach(t => pc.addTrack(t, dest.stream));

// 4. (Optional, paid) swap RNNoise for Krisp SDK at runtime async function upgradeToKrisp(apiKey) { const krisp = await import("@krisp-ai/sdk"); await krisp.init({ apiKey }); // replace the AudioWorkletNode in the graph } ```

Pitfalls

• Stacking three suppressors — running WebRTC NS3 + RNNoise + a server-side suppressor over-suppresses speech and erases sibilants. Pick one client-side and at most one server-side.
• Using `getUserMedia` constraints alone — `noiseSuppression: true` triggers NS3, which is weaker than RNNoise. Set it false when you ship a custom worklet.
• Forgetting AEC — RNNoise is a noise suppressor, not an echo canceller. Keep AEC3 on or you will get loop howl on speakerphones.
• Loading RNNoise via `