Salesforce Live Translation¶

Status: Current | Last reviewed: 2026-10-14

Use this page when Salesforce Service Cloud Voice agents need to talk to customers who speak a different language. TrueServe translates both directions of the live call: the agent hears the customer in the agent's language, and the customer hears the agent in theirs.

Live Translation works on calls routed through Amazon Connect or Genesys Cloud.

Contact center administration¶

Administrators open TrueServe Contact Center Settings from the Salesforce App Launcher. The TrueServe Contact Center Administrator permission set and the TrueServeSCVGrant credential grant are required. Choose the contact center, change the desired settings, and select Save settings.

Use Validate connection and refresh lists to retrieve the real provider queues and phone numbers. Use Manage phone channels and queue mappings to set the primary routing Flow and fallback queue for each number. The default Flow and queue on the TrueServe settings page apply only when no phone channel matches. The phone number must already be assigned to the TrueServe inbound flow in Connect. Saving the provider or number does not provision or move AWS resources.

On the native Partner Telephony Contact Center setup page:

• Queue Mappings match a Salesforce queue to a Connect queue. Your routing Flow can choose the Salesforce queue; Connect delivers the call to eligible agents.
• Contact Center Channels associate a phone number with its routing Flow and fallback queue. Different numbers can use different routing behavior.

The previous Button Assignment / Dialing Options sample fields did not control calls. Use the typed admin settings above. Agents continue to save their personal preferences under Phone → Telephony Settings → My call defaults. Administrator restrictions take precedence. Refresh the console to see capability changes; updated defaults apply to future calls.

Map a Salesforce user to Amazon Connect¶

When the Salesforce and Amazon Connect usernames differ, save an explicit mapping before adding the user to the contact center:

1. Open TrueServe Contact Center Settings and choose the Amazon Connect center.
2. Under User mappings, check whether the user already appears with From Salesforce Setup. Those mappings are reused automatically.
3. To prepare a new user, select Include users not yet added, then Add user mapping. By default, searches show current contact-center members.
4. Search Salesforce user by name or username, then search and select the matching Amazon Connect user. Expanded search includes active Salesforce users without a contact center and all users in this Connect instance.
5. Select Save settings. Existing mappings remain alongside the new row.
6. Follow Manage contact center users, then use Add Users for that user. If Add Users previously failed, retry it after saving.

You can edit a new mapping or select Remove mapping, then Save settings. Manage rows marked From Salesforce Setup through Manage contact center users; they reflect the existing membership and cannot be overwritten here. Removing a connector-only row does not delete a Connect account. Saving does not assign Salesforce queues or Connect routing profiles. Use Refresh Connect users to reload the directory after account changes or a lookup error.

The editor requires a selection from each picker and rejects duplicate Salesforce users. Add Users verifies that the Connect user exists in this center's instance. If the page says the settings changed since you opened it, Reload discards your draft and reads the latest settings; reapply your edit before saving.

Mappings take effect on the next Add Users request. No connector restart or Execute Anonymous script is needed. If User mappings is missing, ask your Salesforce administrator to deploy the updated org-local contact-center settings components and Apex controller.

What “We’re adding users to the contact center” means¶

Salesforce queues a background partner-sync job and emails you its result. TrueServe finds and verifies the selected users' existing Connect accounts and returns their identity pairs. Salesforce then completes the contact-center membership and routing mappings. Check the completion email before assuming the users were added successfully.

The Contact Center Users list in Setup is actual membership. A new mapping saved on the TrueServe page prepares the identity lookup; it does not add the user to that list. Once Add Users succeeds, the TrueServe page reads the Setup mapping automatically. No new Connect account is created by this synchronization.

See pending or failed additions¶

Open TrueServe Contact Center Settings → User sync status. It refreshes every 15 seconds without clearing your unsaved edits; Refresh sync status checks now. Requests appear when Salesforce sends them to TrueServe. The earlier Salesforce queue is not visible here.

Failed additions are shown even though they are absent from Setup’s member list. An unmatched user usually needs a mapping because a sandbox Salesforce username differs from the Connect username. Waiting longer does not fix that failure. Recorded connector attempts are retained for 30 days and begin with deployment of this feature; older attempts are not reconstructed. The page does not retry requests automatically.

Receive an inbound call¶

1. Sign in to Salesforce and the Connect phone system.
2. Set Omni-Channel/Phone to the available status used by your contact center.
3. Call the contact center's assigned number from another phone, then accept the incoming call in Salesforce.

You must belong to the Salesforce contact center and queue, and your Connect routing profile must include the mapped queue. An offline agent does not ring; the caller waits in Connect. Adding a user only to Salesforce is insufficient. For queue transfers, the provider also needs a Quick Connect associated with a queue in the agent's routing profile.

What the agent sees¶

The Call controls panel shows visible On/Off switches for the current call. It can sit on a VoiceCall/Case record page or in the Service Console utility bar. Controls become available when a call connects and follow the active phone call. The accent approach and voice appear below each accent switch. Loading and fallback messages show whether the requested processing is ready.

Set saved defaults, accent approach, voices and grammar repair only in Phone → Telephony Settings → My call defaults. The call page has no defaults editor.

Both language pickers appear when translation is on and usually fill themselves in — see Automatic languages. A picker overrides the published translation agent's configured pair for that call only, so one published agent can serve several language pairs.

The pickers show one entry per language, including Swedish. They do not show country-specific variants because Live Translation uses the base language (es, for example) rather than separate Spanish variants.

Changing a language while translation is running restarts it, with a two-to-three second gap in translated audio. That is expected: the speech engines fix the language when they connect.

To start translation automatically on future calls, use Start translation automatically in Phone settings. The call-page toggle changes only the connected call.

Compare Amazon translation modes¶

You can select published AWS translation agents in your Phone defaults or the contact center defaults. Configure the two AWS profiles to compare Polly PCM response streaming with bidirectional streaming while keeping the same voices and language pair. The call controls work the same way for both profiles.

Save your call defaults¶

Open the phone utility and select its Omni-Channel Settings gear. Scroll to Additional Settings → Telephony → My call defaults. This is the only place to edit your saved call preferences, accent approach and voice. Refresh Salesforce after a package upgrade to load the updated settings.

1. Choose whether translation should start automatically, and optionally choose default languages and a translation agent.
2. Set noise cancellation and background voice cancellation for each direction. Choose Krisp or ElevenLabs as the accent approach. Set Start accent conversion automatically to Off for manual activation, or On to enable it on future calls. Choosing a provider or voice alone does not turn it on.
3. For ElevenLabs, select Choose voice. Search by name, accent or description, press Sample to listen, then Use voice. Use organization voice clears your personal voice choice. The microphone voice is what the customer hears; the customer-audio voice is what you hear.
4. Press Save defaults. Changes apply to future calls and follow your Salesforce account across devices. Saving does not interrupt the current call. The phone's separate Save/Cancel buttons apply to outbound queue selection.

Voice search lists voices accessible through your organization's configured ElevenLabs account, including any library voices already added there. Samples play only to you, one at a time. Selecting a voice does not clone a voice or add a new library voice to the account. A sample demonstrates the voice's sound; it does not measure accent-conversion delay or guarantee recognition accuracy.

Organization policy still controls feature availability. If accent conversion becomes unavailable, the call shows a fallback message and keeps working noise cancellation enabled. Unavailable ElevenLabs does not switch your accent approach to Krisp. Check the status message as well as the switch: On means you requested conversion; the status shows whether it is running. If your Phone settings are still loading, the switch keeps your On request while the current audio mode continues. Switching Off cancels the pending start. If another window saves while you are editing, reopen My call defaults in Phone settings before saving again. A failed save leaves the previous defaults in place.

On a connected call, use Call controls → Agent microphone → Accent conversion to enable or disable the phone-selected voice for the customer. Use the customer audio switch to change what you hear. Each call captures your saved Phone profile when it connects, and its switches keep using that profile. These switches never save defaults. A provider, voice or grammar change saved during a call takes effect on the next call.

Call audio cleanup and Accent Conversion¶

The Call controls panel has two audio sections:

Turning both Customer audio switches Off restores the original call audio directly. Your microphone settings are independent, and hold still silences customer playback. If customer audio is missing, these switches provide a direct bypass for checking it.

The connector reports browser playback failures and automatically checks incoming audio levels and playback state during calls. Support can use those diagnostics without asking you to record a call or copy console output. The checks send level measurements, not audio recordings.

Choose Krisp or ElevenLabs¶

In Phone → Telephony Settings → My call defaults, use Accent approach to choose Krisp or ElevenLabs (experimental). Save before the next call. The microphone and customer-audio choices are independent. Your administrator must configure the ElevenLabs tenant profile and enable its connector runtime. Krisp remains the organization default. During the call, use the page's Accent conversion On/Off switch; the provider and voice stay as selected in Phone settings. Krisp may continue to provide noise cancellation while ElevenLabs handles accent conversion.

When the profile permits it, Phone settings also offer Grammar repair for each direction. The first phrase bypasses repair to prioritize startup. Later phrases use bounded repair and retain recognized wording when an edit is late or fails the preservation checks. Grammar does not reliably fix misheard names or facts.

ElevenLabs uses an English synthesized voice and adds a phrase delay. Its warm browser retests started output in 1.77–2.08 seconds, with about 2.1 seconds of additional connection warmup. Live call timings depend on speech and network conditions. Mute and hold discard queued speech; resume starts a fresh stream and passes the original voice while it reconnects. The connector displays loading or fallback when the requested mode is unavailable. Translation receives original audio through its own branch.

Playback buffering adjusts automatically when ElevenLabs audio arrives unevenly. It gathers enough audio before restarting after a shortage, which can replace repeated choppy gaps with a longer recovery pause. The usual initial reserve is 600 ms; the adaptive target can grow to 1.2 seconds. That target is only the playback buffer, not the total speech delay. Short final phrases still play without needing to fill it. No additional setting is required. This cannot remove pauses already present in the generated voice or guarantee startup within two seconds on every call.

Administrators configure voice, model, buffering and repair in Accent normalization settings.

Noise Cancellation starts from the organization setting and is normally on in both directions. Background Voice Cancellation applies to the agent microphone only and turns on automatically when the selected device and sample rate support it. Accent Conversion is normally off. Switching a toggle requests the change for the current call. The page switches are disabled between calls. While a new model loads, the previous audio mode continues; the status shows loading until the new mode is ready. Routine organization-setting refreshes do not restart working audio enhancement.

Current-call changes reset to your saved defaults when the call ends. Fields set to Organization default inherit administrator policy. The connector enforces this even when the utility panel is closed and restores the live state when you reopen it.

These settings also apply when Live Translation is off. When inbound Accent Conversion is on, the agent hears the converted customer audio while Live Translation receives a separate noise-cleaned copy of the original accent. The customer never hears a duplicate audio stream. If the browser, model, device, sample rate, or CPU cannot run a requested feature, the call continues on the unprocessed path and the setting can be retried on the next call.

Check noise cancellation¶

For Agent microphone, listen from the customer end of a test call. Your headset's own microphone monitoring is not evidence of what the customer receives.

1. Leave Translate this call, Accent conversion, and Background Voice Cancellation off to isolate noise cancellation.
2. Make a repeatable noise near the selected microphone, such as a fan or typing. Speak a short phrase, then leave a pause so the noise can be heard on its own.
3. Under Agent microphone, compare Noise Cancellation Off, On, then Off. Wait for Audio enhancement ready after each change. The customer should hear less noise with On while your speech remains audible.
4. Restore the desired live switches afterward. These switches do not save Phone defaults. Repeat the check after changing microphone or headset.

To test Customer audio, make the noise at the customer end and compare that section's switch while listening through your headset. Other people speaking in the background are a separate Background Voice Cancellation test.

Test accent conversion without translation¶

Your administrator must enable the chosen accent provider. In Phone → Telephony Settings → My call defaults, select the accent approach and voice, set Start translation automatically → Off, and save before the call. Refresh Salesforce after a package upgrade.

1. Place or accept a call, open the page's Call controls, and leave Translate this call off. Translation should show Idle.
2. Enable Accent conversion under Agent microphone to change the voice the customer hears. Enable it under Customer audio to change the voice the agent hears. Test each direction separately first. With only the microphone's accent switch On, have the customer speak too: you should still hear their normal voice through the selected headset or speaker.
3. Wait for Audio enhancement ready, then compare the same spoken phrase with Accent conversion on and off. No language-picker change is needed.
4. Keep the call running for at least a minute and try mute/unmute and hold/resume. Enhancement should remain steady while the call is active; muted speech should not reach the customer.

Using standard call audio means a requested enhancement is not running. Check the displayed reason; a disabled organization policy cannot be overridden by an agent's toggle. End the call when finished: the per-call audio choices reset.

The accent controls process the original voice. If you later enable translation, the translated speech uses the translation agent's synthesized voice.

Accent normalization examples¶

Open /accent-normalization/comparison.html on your connector's base URL. The Accent Normalization page uses Deloitte Agent Studio branding and shows seven rows: female Krisp clean and Office Conversation background pairs, the supplied Sanas demo pair, two ElevenLabs fast startup comparisons, and two earlier buffered test comparisons. The fast-startup rows have Original, Phrase buffering, and With grammar repair players. The grammar version uses Gemini 3.5 Flash-Lite after the first phrase. Earlier rows retain Original, Buffered voice, and With grammar repair. These recordings use the male colleague's speech as input and an American male synthesized voice as output, without translation. The former male Krisp examples are removed.

Starting a recording pauses the others. Individual downloads are MP3 files, and Download All Examples includes eighteen mono recordings and ten stereo timing comparisons, plus transcripts, cut timestamps, source details, and measurements. On smaller screens, the three-player rows stack vertically.

Under each processed Krisp or ElevenLabs recording, expand Timing · original L / output R. Use stereo headphones: the original plays in the left ear and the processed output in the right, preserving the delay. Each timing player also has Download timing MP3. Only one recording plays at a time, including when switching between mono and timing players.

The fast-startup mono and stereo clips retain the entire startup pause. Their Time to first voice labels measure from the first spoken word: 1.89 seconds for both clean outputs, and 1.78 seconds without grammar versus 1.80 seconds with grammar for office audio. These are test playback timings with the providers already connected, not live-call measurements or the delay of every later word. The stereo clips keep both channels on the recorded timeline; no channel is shifted to make the output seem faster.

The Krisp timing examples use fresh streaming captures, shared with their mono examples. Their approximate audio delay is 270 ms in these tests. The earlier ElevenLabs rows show median word delay from their streaming test with simulated playback buffering. Different recordings and measurement methods make this an illustrative comparison, not a live Salesforce/Connect latency measurement. Sanas retains its supplied demo excerpts without a timing player.

The earlier standalone ElevenLabs Buffered voice and With grammar repair examples start promptly, with their initial silence removed. Speech and internal pauses are unchanged. Their stereo timing versions retain the full startup wait, including the extra 1.5-second playback buffer; use those to judge delay. Individual downloads and the ZIP use the same versions as their corresponding players. Grammar repair uses early transcript clauses; incomplete clauses and rejected edits retain the original wording. These recordings still exceed the one-second delay target and can retain misrecognized words. The new grammar tests kept startup below two seconds but did not materially improve wording over phrase buffering alone; misheard words can remain. Within each Krisp or ElevenLabs row, the output comes from the same prepared source as the original. The Sanas excerpts have different wording and background conditions, so they are listening references rather than a controlled comparison with the other technologies.

A separate Live Voice Relay experiment compares an American male ElevenLabs voice with and without optional grammar repair. The updated comparison is at /krisp/listening-tests/elevenlabs-live-relay-2026-09-29-v2/streaming.html on the same translator host. Compare Earlier transcript, Buffered continuous speech, and Buffered speech + early grammar repair. Expand Compare adding the buffer alone to hear its effect without regenerating the earlier voice.

The buffered versions add a 1.5-second startup wait to smooth playback; the recordings preserve that wait. Open Hear original and output together for a timing MP3 with the original on the left and generated voice on the right. On that separate experiment page, Download All Examples includes eighteen MP3s and measurement details. Review grammar changes shows the wording that was retained or adjusted. Incomplete clauses and rejected edits use the original text.

These experimental paths still exceed the one-second delay target; grammar repair also retained a misrecognized phrase. The recordings are mostly grammatical, so the repair step made only a small wording change. These are listening examples with simulated playback timing and do not change the voice or settings of your live Salesforce calls.

Automatic languages¶

Both fall back safely. An unrecognized locale or a blank/misconfigured VoiceCall field just leaves the administrator's default language in place, so the worst case is "translated with the default pair", never "no translation". If an agent picks a language by hand, their choice wins for the rest of that call.

During a call¶

• Turn Translate this call on while the call is connected — or before it, and it waits. If your administrator turned on automatic start, translation begins by itself on every call. The customer's speech reaches the agent translated; the agent's speech reaches the customer translated in a synthesized voice.
• Translation stops by itself when the call ends, and the translation pipeline shuts down with it. There is nothing to remember to turn off.
• Mute and hold work normally — translation stops sending the agent's voice while muted, and pauses entirely while on hold.
• Turning Translate this call off returns the call to normal untranslated audio. The agent's own microphone is restored, so you can stop and restart mid-call.
• Use a headset. Translated audio played through open speakers can be picked up by the microphone and re-translated.

Both sides of the conversation appear in the Salesforce transcript, each turn twice: the words as they were spoken, and the translation. In the native Salesforce transcript the translated line is prefixed [Translation]. In a TrueServe Conversation component the line under each message says which it is — Caller and Caller Translation, Agent and Agent Translation — alongside the time, with Call started and Call ended rows at either end. See Salesforce components.

While translation is running it is the only thing writing the transcript, so a translated call never shows a duplicate set of turns from Amazon Connect's own transcription.

You hear the other person quietly underneath¶

Each side also hears the original voice at a low level beneath the translation — about a tenth of normal volume — dropping almost to silence while the translated voice is speaking and coming back up in the gaps. This is deliberate. It keeps the tone, urgency and turn-taking of a real conversation, and stops the line sounding dead between translated phrases.

It works the same way in both directions, so the customer hears the agent faintly too. If it feels too loud or too quiet, an administrator can change the level per direction on the translation agent: open the translation flow and use Echo Volume under Echo Settings (5–50%, in steps of 5). Each direction is set separately, so the agent's level and the customer's level are independent.

Two things to know about that setting:

• The slider shows 25% before you set it, but a flow that has never been adjusted actually runs at 10%. Move the slider and save once if you want the displayed number to be the real one.
• The Echo Original Audio switch above it does not currently turn the quiet original voice off. To make it as quiet as possible, set Echo Volume to its minimum instead.

Muting still mutes completely — the customer hears nothing at all, not even the quiet original.

Administrator setup¶

1. Create and publish a translation agent¶

In TrueServe:

1. Create an agent with type Speech to Speech and subtype Translation.
2. Configure both flows — incoming (customer → agent) and outgoing (agent → customer) — with the speech services and voices you want, and a language pair as the default.
3. Publish the agent. A draft will not work.
4. Copy the agent's id.

Both directions must be configured. If only one flow is set up, translation fails to start rather than working in one direction.

To use Krisp Voice Translation, choose Krisp → Voice Translation for the translation stage in each direction. Select the source and target BCP-47 languages, male or female translated voice, optional vocabulary terms and dictionary pairs, and which audio/transcript/text outputs the flow should emit. Background Voice Cancellation in this stage is Krisp's server-side option for that translation direction; on an incoming flow it cleans translated audio only and does not change the quieter original customer audio underneath it.

The language pair you set here is the default. When the panel supplies languages — from the agent's Locale, the VoiceCall field, or a picker — those win for that call, so one published agent covers every pair its speech services support. Set the pair you expect most often.

2. Point the connector at that agent¶

Choose one, in order of precedence — the first one set wins:

For the org-wide setting, edit the Call Center definition and add:

<item sortOrder="9" name="reqTranslationAgentId" label="Translation Agent Id">3123</item>

Replace 3123 with your published agent's id.

If none of the three is set, agents see an error when they enable Translate this call instead of silent failure.

3. Place the panel¶

In Lightning App Builder, add TrueServe Call Settings in either location:

• Utility bar: app-wide controls that follow the active connector call.
• Record page: switches follow the active phone call. Automatic language lookup is scoped to that VoiceCall or a record linked through VoiceCall.RelatedRecordId.

Both placements use the same live status channel and support the same optional properties:

For fully hands-off operation: tick Start translation automatically on each call, set Customer Language VoiceCall Field, and leave locale derivation on. Agents then never open the panel — the languages resolve per call and translation starts and stops with the call.

The customer-language field can be any Text field on VoiceCall holding a language code such as es. Locale-shaped values such as es-ES, es_MX, or sv-SE also work and are reduced to their base language. Create it in Setup → Object Manager → Voice Call, and populate it from your routing flow or IVR.

Troubleshooting¶

See also¶

• CCaaS routing to agents
• Adding an integration
• Creating an agent