❓ Questions / Help / Support How to prepare data?

[banjamn]

❓ Questions and Help

We have a wiki available for our users. Please make sure you have checked it out first.

1、How many hours of data do I need to prepare to conduct effective fine-tuning?
2、When conducting fine-tuning, is it necessary to use data from the public dataset? If so, what is the ratio of public data to private data used in the training process?

[snakers4]

Hi,

1、How many hours of data do I need to prepare to conduct effective fine-tuning?

About 10 - 100 hours should be enogh.

2、When conducting fine-tuning, is it necessary to use data from the public dataset? If so, what is the ratio of public data to private data used in the training process?

It would heavily depend on how different your target domain is from the rest of the data.

[banjamn]

@snakers4 Thanks. My data is normal speech data, and I just want to try fine-tuning. What proportion would be reasonable to start with?

[snakers4]

It would make sense to start just with the domain data and track the performance. Then add public data progressively.

[banjamn]

great! Thank you very much!

[banjamn]

@snakers4 I couldn’t find “Common Voice Corpus 16.1” at https://datacollective.mozillafoundation.org/datasets?q=Common+Voice+Corpus. Has it been changed?