ICACT20260217 Question.1
Questioner: maho.hoshiai@gmail.com 2026-02-16 오후 3:12:39 |
ICACT20260217 Answer.1
Answer by Auhor wtlim@etri.re.kr 2026-02-16 오후 3:12:39
Chrome  Click!! |
| How does the VAD threshold affect WER? |
Thank you for the question.
There is a trade-off between the VAD threshold and WER.
When the threshold is higher, VAD becomes more aggressive and removes more segments, which can degrade ASR performance.
When the threshold is lower, VAD keeps more segments, which preserves more speech information and can improve ASR performance. |
ICACT20260217 Question.2
Questioner: maho.hoshiai@gmail.com 2026-02-16 오후 3:06:52 |
ICACT20260217 Answer.2
Answer by Auhor wtlim@etri.re.kr 2026-02-16 오후 3:06:52
Chrome Click!! |
| Have you tested this framework on tasks other than ASR? |
So far, we have mainly evaluated the framework on ASR tasks.
We have not yet conducted extensive experiments on other downstream tasks,
but we believe the framework can be extended to other tasks as future work. |
ICACT20260217 Question.3
Questioner: gonghao23@mails.ucas.ac.cn 2026-02-23 오후 11:14:07 |
ICACT20260217 Answer.3
Answer by Auhor wtlim@etri.re.kr 2026-02-23 오후 11:14:07
Chrome Click!! |
| Generalization of the Post-Enhancement Module: The enhancement network in this framework is trained separately for each codec (Lyra and DAC). Would a single, codec-agnostic post-enhancement model trained on mixed codec distortions achieve comparable WER improvements, and what trade-offs might arise in terms of model complexity and recognition robustness? |
|
ICACT20260217 Question.4
Questioner: gonghao23@mails.ucas.ac.cn 2026-02-23 오후 11:14:22 |
ICACT20260217 Answer.4
Answer by Auhor wtlim@etri.re.kr 2026-02-23 오후 11:14:22
Chrome Click!! |
| VAD-Induced Boundary Discontinuity: The paper acknowledges that removing non-speech frames via Silero-VAD introduces slight boundary discontinuities, which causes a marginal WER increase (e.g., from 6.70% to 7.38% for DAC). How might the framework be redesigned — for instance, by incorporating smooth boundary transitions or context-aware silence padding — to mitigate this degradation before post-enhancement is applied? |
|
