ICASSP 2026 Human-like Spoken Dialogue Systems Challenge

General Information

1. Q: Unable to register using Google Form.

A: If registration via Google Form fails, users in mainland China can also sign up using the following link: [Tencent Docs] ICASSP 2026 HumDial Challenge Registration https://docs.qq.com/form/page/DY2tvT3FvWnRMZXRp

2. Q: Are there any restrictions on the type and size of base models that can be used for the competition?

A: There are no restrictions on the type or size of the models. Participants are free to use any open-source, pre-trained model.

3. Q: Are there restrictions on the tools used for audio synthesis? Specifically, are we required to use only open-source TTS models and prohibited from using commercial APIs like Doubao?

A: Yes, that is correct. Audio synthesis must be performed using open-source TTS models only. The use of commercial APIs is not permitted.

4. Q: For Track 1, is a purely cascaded structure (ASR + LLM + TTS) allowed?

A: No, it is not allowed. You can only use a “thinker-talker” structure or an end-to-end structure.

5. Q: For the 3 tasks in Track 1, can we use a separate model for each?

A: It must be a single dialogue model.