Two-Way Noise Cancellation

You can engineer your agent's environment; you cannot engineer your caller's. That asymmetry defines the problem. On the agent side you have a written audio standard, floor policy, and software — three overlapping controls. On the caller side you have exactly one control: software that cleans the inbound signal after it arrives. There is no policy lever, no coaching lever, no scheduling lever for a customer phoning from a moving bus.

Yet procurement conversations run 90% on the outbound half. We think that's an artifact of who complains: clients and QA hear the agent's noise, so it gets a line item; agents absorb the caller's noise privately, so it becomes invisible fatigue. Ask your agents which direction bothers them more. We have yet to hear "my own background."

Customer-side noise cancellation runs on the agent's machine, in the same software layer as outbound cleanup. The received audio is processed the instant before the agent hears it: non-speech noise is stripped, the caller's voice is lifted out of the mess, and the agent's headset gets the cleaned version. Nothing changes for the caller and nothing touches your telephony — it's the same virtual-device model described in our main ops guide.

On Krisp's published call-center plans this is a distinct line item — "Customer Noise Cancellation: removes background noises coming from customer's side" — included in the entry CC Core tier at $15 per agent/month billed annually, alongside agent-side cancellation and voice isolation (krisp.ai pricing page, verified 2026-09-20). The related listener-side capability for accents — adapting incoming speech so agents comprehend faster — is covered in our accent clarity guide; the two features solve neighbouring halves of the same comprehension problem.

Inbound cleanup pays in small, compounding denominations:

• Repeat-request loops. The "could you spell that?" / "was that B or D?" cycle is pure handle time. Cleaning the inbound signal attacks it directly — we put numbers on this in background noise and handle time.
• Data-capture errors. Misheard digits become wrong addresses, failed verifications and callbacks. A callback is the most expensive form of repeat-request: it costs a whole second call.
• Verification friction. Security questions asked through traffic noise get asked twice — awkward on a normal queue, genuinely risky on a fraud-sensitive one.
• Agent fatigue. Decoding degraded speech for a full shift is cognitive work. Agents describe end-of-day "listening exhaustion"; cleaning inbound audio is the only lever that touches it, because the fatigue source is the caller's environment, not the agent's.

Inbound cleanup is not evenly valuable across programs. It pays fastest where callers are structurally noisy: roadside assistance and logistics (cabs, depots, wind), field-service support (job sites), travel disruption lines (airports, announcements every forty seconds), and any queue that spikes during commutes. It pays slowest on queues where callers dial from quiet desks. Before you evaluate, pull a day of recordings from your own noisiest program and count the agent-side repeat-requests — that number, not a category average, tells you which of your queues is silently paying the inbound tax and should host the pilot.

Two-way tools ship as one install, so you can't A/B the directions separately in production — but you can instrument the inbound half. In your pilot (design details in the ops guide), have QA tag three things per scored call for two weeks before and two weeks after: repeat-requests by the agent (inbound problem), repeat-requests by the caller (outbound problem), and data-capture corrections. If the inbound-tagged numbers don't move, you've learned your callers are quieter than average and the outbound half must justify the seat price alone. Either way, you decide on your data — the metrics guide has the full tagging scheme.