Alibaba's Qwen team has released Qwen3.8-LiveTranslate, a real-time simultaneous interpretation model that listens to live speech with optional video frames and returns translated text and speech while the speaker is still talking. The model cuts average lag from 2.8 seconds to 2.3 seconds, a roughly 18% improvement in responsiveness. This represents a meaningful step forward for multimodal AI systems that must balance speed with accuracy in live translation scenarios.

How Does Real-Time Translation Actually Work?

Simultaneous interpretation has always been a tradeoff. Waiting longer gives the model more context to produce accurate translations, but speaking sooner reduces delay for the listener. Qwen3.8-LiveTranslate rebuilds this loop with a new Interleave architecture that improves faithfulness, fluency, and conciseness without sacrificing speed. The model measures performance using a metric called Length-Adaptive Average Lagging (LAAL), which tracks how far the translation trails the source speech on average while avoiding systems that over-generate output.

The system accepts audio and optional images as inputs and outputs both text and audio. Visual cues such as lip movements, gestures, and on-screen text help the model understand speech in noisy rooms and resolve ambiguous words. The documentation recommends sending no more than two images per second to maintain efficiency.

What Languages and Features Does It Support?

The model understands 60 languages total, though it can speak only 29 of them with audio output. The remaining 31 languages return text-only translations. Speech output covers major languages including Chinese, English, Arabic, German, French, Spanish, Japanese, Korean, and Hindi.

Beyond basic translation, Qwen3.8-LiveTranslate includes several features designed for professional use:

• Real-time speaker diarization: The model distinguishes between multiple speakers in multi-party speech and preserves each speaker's voice through stable voice cloning, with modes that re-clone before each response for multi-speaker sessions.
• Synchronized bilingual display: Source text and translation appear on screen together, with source transcription streaming as separate events alongside the translation stream.
• Long-context disambiguation: The model uses conversation history to resolve names and terminology, ensuring that a name introduced early in a meeting stays consistent throughout the translation.
• Hotword configuration: Teams can map source terms to fixed target translations, with documentation recommending no more than 1,000 hotwords per session.

How Can Teams Access and Deploy This Model?

The model is deployable as a hosted API through Alibaba Cloud Model Studio and QwenCloud under the model ID qwen3.8-livetranslate-flash-realtime. Developers connect via WebSocket Realtime API with continuous audio streaming and server-generated responses. Default audio input is 16 kilohertz PCM, with 24 kilohertz PCM output. The default voice is named Tina, though clients can customize output modalities to text only or text and audio.

Pricing varies by region. In Singapore, audio input costs $7.50 per hour, image input costs $0.55 per image, text output costs $20 per million tokens, and audio output costs $30 per million tokens. Beijing pricing is lower, at $5.65, $0.47, $14.13, and $22.61 in USD respectively. One hour of continuous speech in and speech out costs approximately $1.54 in Singapore before accounting for text and image tokens. The context window is 53,248 tokens total, with 49,152 reserved for input and 4,096 for output.

The model builds on Alibaba's Qwen-Omni stack, large-scale multimodal data, and cross-language and cross-modal alignment techniques. It represents the real-time version of Qwen3.8-LiveTranslate-Flash, which also supports offline audio and video translation for teams that do not need live interpretation.

What Are the Practical Implications for Global Teams?

For organizations conducting international meetings, conferences, or live events, reducing translation lag from 2.8 seconds to 2.3 seconds improves the natural flow of conversation. This improvement matters most in scenarios where participants need to respond quickly or where pauses feel awkward. The addition of speaker diarization and bilingual display makes the system practical for multi-party conversations where tracking who said what is critical.

The API-only deployment model means teams do not need to manage infrastructure themselves, though they do need to integrate with Alibaba's cloud services. Default rate limits are 10 requests and 100,000 tokens per minute, which should accommodate most meeting scenarios. Teams should note that function calling, structured outputs, batch inference, and fine-tuning are not currently supported on the platform.