General Questions

Frequently asked questions will be listed below.

Q1. Can participants join both tracks?

Each team can participate in only one track — either TTS or SASV. Participants cannot join both tracks at the same time.

Q2. When will the test data for evaluation be released?

For the SASV track, the test dataset will be released after November 15th, once we receive outputs from the TTS track teams.

Q3. There seems to be an error in Table 1 of the paper regarding TITW-Easy and TITW-Hard sample counts. Which one is correct?

Yes, there is a typo in Table 1 of the paper. The numbers of samples for TITW-Easy and TITW-Hard are reversed. The correct statistics are:

Set	samples	Avg dur (s)	Tot dur (h)	Avg # words
TITW-Easy	248,024	2.51	173	10.55
TITW-Hard	282,606	2.42	189	10.84

For TTS Track

Q4. Are pretrained datasets and pretrained models allowed on the TTS track?

Yes, pretrained datasets and models are allowed. However, the final submitted model must be fine-tuned on the TITW dataset in the last phase.

Q5. Is it allowed to combine multiple independently trained models during inference (e.g., averaging scores or using stacking)?

Yes, combining multiple models during inference is allowed.

Q6. Which evaluation metrics will be used for TTS track.

We will use Versa as our evaluation tool. WER, UTMOS, DNSMOS, and speaker similarity will be calculated using this tool automatically.

For SASV Track

Q7. Is it allowed to use external datasets for the SASV track?

Only the VoxCeleb2 dataset can be used, as it does not overlap with the SpoofCeleb test set. No other external datasets are permitted for the SASV track.

Q8. Are SSL (Self-Supervised Learning) models allowed?

Yes, SSL models are allowed.

Q9. Can SSL models be pre-trained on public datasets other than SpoofCeleb and VoxCeleb2?

Yes, SSL models pre-trained on other public datasets are permitted.

Q10. Are there restrictions on the type of SSL models that can be used?

No, any model that applies self-supervised learning (SSL) is allowed, including but not limited to wav2vec 2.0, HuBERT, and other SSL-based models.

Q11. Can we use foundation models from other modalities, such as large language models (LLMs) for text processing?

Yes, the use of multimodal foundation models, including LLMs, is allowed.

Q12. Can MUSAN/RIR be used for data augmentation?

Yes, the use of MUSAN and RIR is permitted for data augmentation.

Q13. Can we use other data augmentation methods beyond MUSAN/RIR (e.g., TTS-based data generation)?

Yes, it’s acceptable as long as no additional external datasets are introduced.

Q14. How will the evaluation score be calculated?

Each test utterance will be paired with one registration utterance, following the provided evaluation protocol format (a list of <enroll test_trial > pairs).

Q15. Where can I download the evaluation data?

A link has been sent to all registered teams. Full evaluation data will be made public after the challenge closes. If you haven’t received the link, please check the trash folder, or contact Yihan Wu (yihanwu@ruc.edu.cn).

Contact

If you have any other questions, please email the organizing committee: Yihan Wu (yihanwu@ruc.edu.cn).