Compare AWS voice translation modes

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

Create two translation agents to compare Amazon Transcribe, Translate and Polly in Salesforce Live Translation. Each agent provides translation in both call directions.

Configure the first agent

  1. Create a translation agent in Agent Studio.
  2. In each incoming and outgoing flow, add a Speech-to-Speech Translation stage and select AWS → Voice Translation.
  3. Select your Amazon integration. Your administrator enables its Translation capability and grants Transcribe, Translate and Polly access. Installations using a workload IAM role can leave the integration unset.
  4. Open the stage's Properties. For each direction, set Source language and Target language to the speaker and listener locales.
  5. Set Polly streaming mode to Phase 1 — PCM response streaming.
  6. Keep Polly engine set to Generative. Set AWS region and choose a Polly voice supported in that region for the target language.
  7. Save and publish the agent with a name such as AWS Translation — Phase 1.

For a Spanish-speaking customer and English-speaking agent:

Flow Source language Target language Example Generative voice
Incoming, heard by the agent es-US en-US Danielle
Outgoing, heard by the customer en-US es-US Lupe

These example voices were validated in us-east-1. You can use auto to select an available voice, but pin a voice in each direction for a controlled comparison. The stage has Properties only; no translation prompt is needed.

Configure the second agent

Duplicate the first agent and name it AWS Translation — Phase 2. In both flows, change only Polly streaming mode to Phase 2 — Bidirectional streaming, then publish it.

Keep Generative, the region, voices, languages and timing settings identical. Leave Flush each translated phrase enabled for the initial comparison. Phase 2 requires Generative. Phase 1 also supports Neural and Standard, but changing the engine would compare voice models as well as streaming modes.

Compare in Salesforce

  1. In Phone → Telephony Settings → My call defaults, select the Phase 1 translation agent. If personal overrides are restricted, your administrator selects it in TrueServe Contact Center Settings instead.
  2. Start a test call and turn on Translate this call.
  3. Read the same short phrases from both sides, with the same pauses. Check the source transcript, translation meaning, speech completeness and delay.
  4. End the call, select the Phase 2 agent, and repeat.
  5. Alternate the order over several calls. Keep the input devices, network, languages and enhancement settings the same.

The language pickers override the configured pair for that call. An explicitly selected Polly voice must support the resulting target language. Use auto when trying other language pairs, then pin the selected voices for comparison.

The engineering console reports [AWSTranslationPerformance] events for each direction. Those timings help distinguish recognition, translation and synthesis delay; they do not measure the full delay heard by the listener. Ask an engineer to use the replay and timing guide for saved measurements.

Adjust phrase timing

Transcript stability controls when recognition words become stable. Stable phrase duration (ms) allows a stable prefix to be translated before a long utterance ends. A lower duration can speak sooner but gives the translator less context. Neither setting guarantees a maximum delay.

Maximum pending translation age (ms) ends translation if queued work falls too far behind. SCV restores the original call audio when translation fails. An unavailable voice, unsupported language/engine/region combination, or missing AWS permission also produces a failure instead of changing the engine silently.

See also