Accent normalization

Status: Current | Last reviewed: 2026-10-01

Tenant administrators can enable an experimental ElevenLabs alternative to Krisp for Salesforce call audio. Open Settings → Tenant Settings → Accent normalization. This is a tenant audio profile; it does not create a conversational agent or change your translation agents.

  1. Select an active ElevenLabs integration and an English voice ID.
  2. Choose Flash v2.5 or Flash v2. The default profile uses Flash v2.5, English, and speech speed 1.1.
  3. Keep the initial timing defaults: 150 ms transcript stability, 650 ms phrase wait, 600 ms startup audio buffer and 6000 ms maximum queued audio.
  4. To offer grammar fixes, select a Google AI integration and enable Allow optional grammar repair. Agents still choose it separately for each call direction. The first phrase always preserves the recognized wording for faster startup; later edits are accepted only within the available playback budget and preservation checks.
  5. Enable the profile and select Save profile.

Your connector administrator must also enable the SCV runtime for this tenant and Salesforce org. Until both sides are configured, ElevenLabs does not appear as an available choice in Salesforce. Provider secrets remain in Integrations; there are no API-key fields on this page.

The current grammar default is Gemini 3.5 Flash-Lite. Model availability depends on the integration: Gemini 2.5 Flash-Lite returned a model-not-found response in the tested account. An unavailable model, slow repair or rejected edit retains the original recognized text.

These controls trade waiting time against continuity. A smaller startup buffer can start sooner but allows more gaps between phrases. The maximum buffer bounds how far playback can fall behind; exceeding it returns the call to standard audio. First-use provider warmup is separate from speech startup. The implementation's four warm browser retests measured 1.77–2.08 seconds from source speech to output speech, plus about 2.1 seconds to initialize the providers. These are short fixture tests, not a live-call latency guarantee.

ElevenLabs synthesizes a different voice and can repeat a recognition mistake. Grammar repair did not establish a wording or startup improvement in these examples. Compare both modes with representative calls before adopting a profile broadly. Muting or holding clears queued speech; resume warms a fresh stream, passing the original voice during reconnect, and does not replay words from before the pause.

For deployment, create a Tenant Token with only Accent Runtime (accent:runtime) under the API-token controls and give it to the trusted SCV backend. This scope can resolve the enabled profile's provider credentials, so keep the token server-side. Ordinary configuration and personal tokens do not substitute for it. Connector operators use the handbook’s SCV deployment instructions.

Agents choose their provider in Salesforce Call Settings.