Fixed #9 (Preprocessors do not have access to the user_id and sentence) #10
Conversation
|
See issue #9 (Preprocessors do not have access to the user_id and sentence). |
|
OK - got it. But how would the preprocessor/user get the sample path with a given user_id and sentence practically? |
|
Common Voice will have a alpha release. This alpha release will contain audio + clips.tsv. Contributors will download the alpha release and then write a preprocessor for their language for this alpha release. So preprocessor writers will have the alpha release with audio on their hard drive and be able to hear the audio of a particular (user_id, sentence) pairing when the write the preprocessor by looking in the tsv for the (user_id, sentence) pairing and finding the corresponding audio clip. Yes this ties all preprocessors to the "already recorded" use-case. That is the problem at hand. As issue #9 (Preprocessors do not have access to the user_id and sentence) illustrates. It is impossible to always correctly handle the "not recorded" use-case as we never know how to convert "I am in room 2049" to words without hearing the audio. |
|
In the "not yet recorded" case we would just define how to speak "I am in room 2049". |
|
Yes, but then we have to ask Common Voice (again) to not put digits in to the sentences to be read. Hopefully they'll listen this time 😄 |
|
In my understanding the preprocessors should also be used to preprocess sentences before the actual recording - if not even before merging new sentences into the sentence collection. |
|
They could even introduce a commit-hook that prohibits committing sentences with numbers. |
|
This is only to be used after the fact to create corpora, i.e. this is used once everything is recorded. |
|
Yes they could introduce a commit-hook, but they haven't |
No description provided.