Speech-to-speech

Migrate from closed-source models

Map your current OpenAI Realtime or Gemini Live setup to an equivalent Bailian model.

S2S (speech-to-speech) vs. pipeline

There are two approaches to building voice applications:

• Use S2S for low latency, audio-aware responses, and interactive conversation.

• Use a pipeline when you need voice customization or want to select ASR, LLM, and TTS models independently.

This page covers the S2S single-model approach (Omni and Livetranslate series). For the pipeline approach, select each component separately:

• ASR (speech recognition): Speech-to-text

• LLM (large language model): Text generation

• TTS (text-to-speech): Speech synthesis

Real-time or file mode?

• Real-time (WebSocket): Voice assistants, call centers, and simultaneous interpretation. Streams audio input and speech output. Model names contain -realtime.

• File mode (HTTP): Higher latency but better quality. Ideal for video dubbing, podcast translation, and offline processing. Also supports function calling, web search, thinking mode, and video context (see Companion capabilities below).

Choose a model by use case (S2S single-model approach)

All use cases below use the S2S single-model approach. For the pipeline approach, use the ASR, LLM, and TTS guides linked above.

Companion capabilities of the S2S single-model approach

Qwen3.5-Omni and Qwen3-Omni provide these capabilities natively. With a pipeline, equivalent functionality comes from individual components (typically the LLM).

Speech translation

The following model series support speech translation:

• Qwen3.5-Livetranslate: 60 languages (29 with audio+text output, 31 text-only). Covers Chinese, English, French, German, Russian, Japanese, Korean, Spanish, Portuguese, Arabic, and more.

• Qwen3-Livetranslate: 18 languages and 5 Chinese dialects (~3 s latency). File mode accepts video input for context-aware translations. 7 languages produce text-only output.

• Qwen3.5-Omni: 29 output languages and 8 Chinese dialects. Strong audio/video understanding and web search. Inject terminology and domain context via system prompt. Real-time and file modes.

• Qwen3-Omni-Flash: 11 output languages and 8 Chinese dialects. Inject terminology and domain context via system prompt. Real-time and file modes, at lower cost.

Recommended models

The table lists the entry-point model in each series. To pin a dated version for regression testing or stability, see All models below.

All models

What's next

API documentation by model series:

• Qwen3.5-Omni and Qwen3-Omni (WebSocket, real-time): Qwen-Omni-Realtime

• Qwen3.5-Omni and Qwen3-Omni (HTTP, file): Non-real-time (Qwen-Omni)

• Qwen3.5-Livetranslate (WebSocket, real-time): Real-time audio and video translation - Qwen

• Qwen3-Livetranslate (HTTP, file): Audio and video file translation (Qwen)

• Qwen-Audio Realtime (WebSocket, real-time voice conversation): Realtime Audio Chat (Qwen-Audio-Realtime)