LiveTranslate 2.3 s latency vs STT plus TTS

Qwen3.8-LiveTranslate, listed on September 18, 2026, uses an interleave architecture and reports average lagging (LAAL) down from 2.8 to 2.3 seconds. Sume does not offer live interpretation; for recorded video you can chain stt_create, a translation step you choose, and tts_create.

What the release note says

The Releasebot Qwen feed records two facts for this model: the interleave architecture and the drop in average lagging (LAAL) from 2.8 s to 2.3 s. A 0.5 second reduction is about 18 percent of the earlier figure. The entry does not say how latency was measured, so compare only against your own tests.

Live versus offline

Live translation is a streaming problem: audio goes in while the speaker is still talking and the delay is what the listener feels. A dubbing job is a batch problem: you have the whole file, so quality and timing matter more than the first-word delay.

Sume's hosted tool list contains speech-to-text (stt_create) and text-to-speech (tts_create), both submitted as paid jobs with an idempotency_key. There is no live interpreter among the documented tools, so a latency comparison with LiveTranslate would not be meaningful.

An offline dub pipeline

Each step is a separate job, so you can inspect the text between steps.

• Transcribe the source audio with stt_create and keep the timed text.
• Translate the text with the model or service you trust, and have a speaker of the target language review it.
• Generate the voice track with tts_create, one job per sentence or paragraph.
• Use jobs_wait to collect results, then place the audio on the video with the timeline tools or Sume's other media tools.
• Check duration drift: translated speech is often longer or shorter than the source.

When to stay with a live model

If the use case is a meeting, a call or a stream, choose a live model and evaluate its latency in your network. If the use case is a published video, the offline route gives you a reviewable transcript and a file you can re-render.