Skip to main content
Callers phone from kitchens, cars and shop floors. A TV in the background, a second person talking or street noise reaches the transcriber as speech, and your agent answers things nobody said to it. Noise cancelling cleans the caller’s audio before speech recognition and turn detection: it keeps the person on the call and suppresses noise and other voices around them. Noise cancelling runs on Fonoster, before your application gets the audio. You turn it on per application, and nothing in your code changes.

Turn it on

Noise cancelling is a setting on the application, off by default. It needs a noise-cancelling provider in your workspace’s catalog. Fonoster Cloud offers ai-coustics; a self-hosted install offers it once the operator adds the provider (see Turn-taking and audio).
  • App: go to Apps, open the app, go to the Call handling tab, and pick a provider under Noise cancelling.
  • Autopilot agent: open the agent, go to the Advanced tab, and pick a provider in the Noise cancelling panel.
Choose Save. The list shows None and every noise-cancelling provider your workspace can use. If it shows None only, no provider is available to the workspace yet.
Each filter in filters has a name and optional options: The dashboard sets the filter with its defaults. Use the SDK or the CLI’s --from-file to change enhancementLevel.

What to expect

  • One speaker. The filter locks onto the main voice and suppresses the rest. That is what you want for a person on a handset. It is wrong for speakerphone calls with several people, for a phone menu that answers your outbound call, and for hold music. Turn it off on applications that take those calls.
  • A little delay. Filtering adds about 30 milliseconds to the caller’s audio.
  • Recordings are not filtered. Call recordings keep the audio as it arrived. Only what the transcriber and your application hear is filtered.
  • Twilio-compatible apps too. The filter runs before the audio reaches your app, whatever the protocol.

Bring your own key

Fonoster Cloud runs the filter on the platform’s ai-coustics account. To run it on your own account, add an ai-coustics key on the Vendor keys page, the same way as a speech or language-model key. Your workspace’s calls then use it. See Use your own vendor keys.

When it can’t run

Noise cancelling never drops a call. If the filter can’t start, for example because the provider rejected the key, or if it fails during a call, the call goes on with the caller’s audio as it arrived. The call’s call record notes it, and the call’s page in the dashboard shows an Unfiltered badge with the reason.

Next steps

Use your own vendor keys

Run the filter on your own ai-coustics account.

Call records

See which calls ran unfiltered, and why.

Detect answering machines

Another per-application setting for the first seconds of a call.

Turn-taking

How clean audio feeds turn detection.