dataset #5
-
|
Hello author, after reading your article, I thought of my current application scenario which involves mapping bone conduction speech signals to air-conducted speech signals. However, the dataset is currently limited. How does the size of the dataset affect the performance of an EBNE network? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 5 replies
-
|
Hi, Our approach can be applied to any body-conduction microphone including bone, in-ear, and throat microphones. If your dataset is limited, you can try to perform a pre-training on simulated data (just as we did by applying a low-pass filter on clean speech with roughly the same characteristics as the mic you're tackling). The size of the finetuning dataset is a question that we are currently exploring. We are planning to record 50 hours of air and body-conducted speech. We believe it will be sufficient with well-adapted hyperparameters. Good luck with your application! |
Beta Was this translation helpful? Give feedback.
-
|
I have a dataset of about 5 hours, do you think the model will converge? |
Beta Was this translation helpful? Give feedback.
-
|
Idk, deep learning is an experimental science, so you should give it a try. With adequate pre-training, it might work. 😉 |
Beta Was this translation helpful? Give feedback.
-
|
i am working on it and hope it can work |
Beta Was this translation helpful? Give feedback.
-
I previously used a mapping method similar to the speech enhancement model DCCRN and found that it could only map up to 4000Hz. It was unable to map frequencies between 4000Hz and 8000Hz. Currently, I am trying your model. |
Beta Was this translation helpful? Give feedback.
-
I have found some datasets data1 But in the country where I am located, it is not possible to download such a large dataset :( You can download this dataset and take a look at its data quality. Interestingly, we have previously discovered the same dataset. data2 |
Beta Was this translation helpful? Give feedback.
Hi,
Our approach can be applied to any body-conduction microphone including bone, in-ear, and throat microphones.
If your dataset is limited, you can try to perform a pre-training on simulated data (just as we did by applying a low-pass filter on clean speech with roughly the same characteristics as the mic you're tackling).
The size of the finetuning dataset is a question that we are currently exploring. We are planning to record 50 hours of air and body-conducted speech. We believe it will be sufficient with well-adapted hyperparameters.
Good luck with your application!