@0xLogicrw: Qwen has released the real-time simultaneous interpretation model Qwen3.8-LiveTranslate. It reduces the average latency…

X AI KOLs Timeline Models
real-time-translation ai-model latency-reduction speaker-diarization streaming-output

Summary

Qwen has released Qwen3.8-LiveTranslate, a real-time simultaneous interpretation model that reduces average latency from 2.8s to 2.3s and includes new features like speaker diarization and long-context disambiguation.

Qwen has released the real-time simultaneous interpretation model Qwen3.8-LiveTranslate. It reduces the average latency per word (LAAL, which measures the average lag time in simultaneous interpretation) from the previous generation's 2.8 seconds to 2.3 seconds, and adds new features like real-time speaker diarization, synchronized output of original text and translation, and long-context disambiguation. The model can now distinguish between different speakers, map translations to specific utterances, and make the translated speech more stably retain each speaker's timbre. The original text and translation will also be returned synchronously in streaming fashion. The model will also combine prior context to judge names, terms, and references, reducing ambiguities that arise from looking only at the current sentence. API pricing is basically unchanged from the previous generation. Alibaba Cloud's price list shows that in the Beijing region, the prices for audio input, image input, text output, and audio output for Qwen3.8-LiveTranslate are exactly the same as for Qwen3.5-LiveTranslate; in the Singapore region, all four prices have even seen a slight decrease.
Original Article
View Cached Full Text

Cached at: 09/19/26, 09:09 PM

Qwen has released the real-time simultaneous interpretation model Qwen3.8-LiveTranslate. It reduces the average latency per word (LAAL, which measures the average lag time in simultaneous interpretation) from the previous generation’s 2.8 seconds to 2.3 seconds, and adds new features like real-time speaker diarization, synchronized output of original text and translation, and long-context disambiguation.

The model can now distinguish between different speakers, map translations to specific utterances, and make the translated speech more stably retain each speaker’s timbre. The original text and translation will also be returned synchronously in streaming fashion. The model will also combine prior context to judge names, terms, and references, reducing ambiguities that arise from looking only at the current sentence.

API pricing is basically unchanged from the previous generation. Alibaba Cloud’s price list shows that in the Beijing region, the prices for audio input, image input, text output, and audio output for Qwen3.8-LiveTranslate are exactly the same as for Qwen3.5-LiveTranslate; in the Singapore region, all four prices have even seen a slight decrease.

Qwen (@Alibaba_Qwen): Meet Qwen3.8-LiveTranslate, Qwen’s next-generation real-time simultaneous interpretation model! 📢

Built on an Interleave architecture, it improves faithfulness, fluency, and conciseness while reducing average lagging (LAAL) from 2.8s to 2.3s across 60 languages.

New

Similar Articles

Qwen3-TTS Technical Report

Papers with Code Trending

The Qwen3-TTS technical report introduces a series of advanced multilingual text-to-speech models with voice cloning and controllable generation, featuring a dual-track LM architecture and specialized tokenizers for low-latency streaming.

Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice

Hugging Face Models Trending

Alibaba's Qwen team releases Qwen3-TTS-12Hz-1.7B-CustomVoice, a powerful text-to-speech model supporting 10 languages with low-latency streaming, instruction-based voice control, and robust contextual understanding.