Alibaba’s Apsara 2026 Audio Models, Explained

Alibaba’s audio announcements around the 2026 Apsara Conference address two different jobs: translating a speaker while they are still talking, and building applications that recognize, generate, converse through or create audio. Qwen3.8-LiveTranslate handles the first; the five-model Qwen-Audio-3.1 lineup covers the second. 314

Qwen3.8-LiveTranslate: interpreting speech in real time

Qwen3.8-LiveTranslate is designed for simultaneous interpretation, producing translated output as speech arrives. Alibaba says its Interleave architecture improves translation fidelity, fluency and brevity while reducing LAAL, a measure of average translation lag, from about 2.8 to 2.3 seconds—a half-second reduction, or roughly 18%. Those quality and speed improvements are company-reported claims, not independent test results. 37

Alibaba also describes features for conversations involving more than one person: real-time speaker separation, a synchronized display of source and translated text, and use of earlier context to resolve ambiguous language. Its Model Studio documentation lists simultaneous interpretation and live translation as uses for the real-time model. 71

The five Qwen-Audio-3.1 models

Qwen describes Audio 3.1 as three upgraded capabilities—ASR, TTS and Realtime—plus two new models, ASR-Next and TTS-Next. They serve distinct purposes: 14

• Qwen-Audio-3.1-ASR turns speech into text, making it the transcription model in the lineup. 1426
• Qwen-Audio-3.1-ASR-Next extends audio understanding beyond transcription. Qwen says it can label speakers and align transcripts with timestamps, as well as analyze emotions and non-speech sounds for tasks such as sound captioning and audio questions. 14
• Qwen-Audio-3.1-TTS synthesizes speech from text. Qwen describes multilingual and dialect support, alongside controls for delivery such as emotion, speed and style. 14
• Qwen-Audio-3.1-Realtime is for interactive voice conversations; Alibaba Cloud lists voice assistants and customer-service conversations among uses for its Realtime offering. 141
• Qwen-Audio-3.1-TTS-Next creates fuller audio scenes rather than speech alone. Alibaba says it can generate dialogue, sound effects and ambience together in a single pass from inputs including text, timestamps and reference audio. 24

TTS-Next is the clearest distinction between ordinary speech synthesis and audio production in this release. Alibaba pitches its script-to-soundscape capability for audiobooks, film and television, podcasts and games. It is an intended creative use, not evidence that a generated scene will meet every production requirement without editing. 324