Speech-to-speech

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 to customize voices or select best-in-class ASR, LLM, and TTS models for each stage.

This page focuses on the S2S single-model approach (the Omni and Livetranslate series). If you choose the pipeline approach, select each component from the corresponding documentation:

• 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): For real-time voice interaction such as voice assistants, call centers, and simultaneous interpretation. Supports streaming audio input and speech output. Model names contain -realtime.

• File mode (HTTP): Trades higher latency for better quality, ideal for video dubbing, podcast translation, and offline content processing. File mode also supports companion capabilities such as function calling, web search, thinking mode, and video context. For details, see Companion capabilities of the S2S single-model approach below.

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

The use cases below all assume the S2S single-model approach. For the pipeline approach, choose components from the ASR, LLM, and TTS guides linked above.

Companion capabilities of the S2S single-model approach

Under the S2S single-model approach, the Qwen3.5-Omni and Qwen3-Omni models provide the following capabilities directly. With a pipeline approach, equivalent functionality must come from individual components (typically the LLM).

Speech translation

The following model series support speech translation:

• Qwen3.5-Livetranslate: Translates between 60 languages, with 29 producing both audio and text output and 31 producing text only. Covers major languages such as Chinese, English, French, German, Russian, Japanese, Korean, Spanish, Portuguese, and Arabic.

• Qwen3-Livetranslate: Supports 18 languages and 5 Chinese dialects with a latency of approximately 3 seconds. In file mode, it uses video input to deliver more accurate, context-aware translations. For 7 of these languages, the output is text-only (no speech).

• Qwen3.5-Omni: Supports 29 output languages and 8 Chinese dialects. Offers strong audio and video understanding and web search. Use a system prompt to inject terminology and domain context. Supports both real-time and file modes.

• Qwen3-Omni-Flash: Supports 11 output languages and 8 Chinese dialects. Use a system prompt to inject terminology and domain context. Supports both real-time and file modes, at lower cost.

Recommended models

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

All models

What's next

After you choose a model, see the corresponding API documentation:

• Qwen3.5-Omni and Qwen3-Omni (WebSocket, real-time): 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 speech and audio-video translation with Qwen

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