Speech Commands v0.01 & v0.02 dataset #996
Conversation
yfyeung
force-pushed
the
speechcommand
branch
2 times, most recently
from
2d94ad4 to
aff5912
Compare
|
Hi, I am building a wake word recipe in icefall for this dataset. |
|
Perhaps the recipe should be called "speech_commands" instead of "speech_commands001", and the version should be provided as a parameter for download/prepare. |
Ok, I will implement this. |
yfyeung
changed the title
In that case, I suggest that we support both v1 and v2. |
Ok, I have implemented the recipe "speech_commands002" locally. I will merge v1 and v2 into one recipe. |
yfyeung
changed the title
|
Thanks! You can look at the VoxCeleb recipe for an example where we support v1 and v2. |
yfyeung
changed the title
|
@desh2608 Hi, I have finished this recipe, supporting Speech Commands v0.01 and v0.02 dataset. |
lhotse/recipes/speechcommands.py
Outdated
| :return: the path to downloaded and extracted directory with data. | ||
| """ | ||
|
|
||
| return _download_speechcommands( |
There was a problem hiding this comment.
Do we really need v1 and v2 to have separate download functions even though they share the same logic? I think it's cleaner to rename _download_speechcommands into download_speechcommands and set the default version to the latest one, and then expose the version argument in the CLI (so that we have one CLI program for this rather than two).
lhotse/recipes/speechcommands.py
Outdated
| ) | ||
|
|
||
|
|
||
| def _prepare_train_valid( |
There was a problem hiding this comment.
could you split this function into separate prepare_train and prepare_valid? I don't think you really need a generator here if you restructure the code, and the way to use this as-is is a little confusing.
There was a problem hiding this comment.
OK, I will split it.
|
@pzelasko I have addressed them, please re check it. |
|
LGTM! |
Speech Commands is an audio dataset of spoken words designed to help train and evaluate keyword spotting systems.
paper: https://arxiv.org/pdf/1804.03209.pdf