Key Highlights
On July 15, Alibaba officially released Qwen-Audio-3.0-Realtime, a voice interaction model designed for real-time conversation scenarios. The new model brings upgrades in four areas: reasoning ability, tool calling, emotional expression, and duplex interaction.
The series comes in two variants:
- Plus: Stronger reasoning for complex dialogue tasks
- Flash: Faster response for low-latency scenarios
Both are available through Alibaba Cloud Bailian (Beijing region).
Function Calling
Qwen-Audio-3.0-Realtime supports function calling. Developers can connect maps, databases, and knowledge bases — the model decides during conversation when to trigger these tools. For example, when a user says "find the next high-speed train to Shanghai Hongqiao station," the model automatically calls the query API and returns results.
This matters more in voice scenarios than text because voice conversations tend to be faster-paced with denser interruptions — manually triggering tools isn't practical.
Emotional Expression and Empathetic Dialogue
The model has several improvements in voice expression:
- Adjusts tone, speed, pitch, and emotion based on conversation context
- Supports non-verbal sounds — laughter, sighs, hesitation
- Can create custom voice profiles from audio samples
According to Alibaba's testing data, the model achieved the best score on VStyle, a bilingual public benchmark for voice style adaptability.
Duplex Communication
Qwen-Audio-3.0-Realtime also performs better in noisy environments. The model can identify the primary speaker and reduce false interruptions from bystanders. In multi-speaker scenarios, it uses context to determine who is speaking and switches between speakers accordingly.
Both models accept audio and text input via WebSocket and stream voice and text output. Developers can choose text-only, audio-only, or mixed interaction modes as needed.
Market and Use Cases
Alibaba is targeting four directions: customer service, education and training, entertainment, and emotional companionship. The Plus variant suits scenarios requiring deeper reasoning (customer service, training), while Flash targets latency-sensitive use cases (interactive entertainment, companionship).
Compared to alternatives, Qwen-Audio-3.0-Realtime stands out for combining function calling with emotional expression — it can call external data in real-time while maintaining natural voice interaction pace. However, it's currently only available in the Beijing region; developers elsewhere will need to wait for expansion.