I started to check audio clips in Russian Scripted Speech tab. And about 40-50% of them from 100-200 I checked were just synthesized voices reading these sentences. Is it health practice and should be accepted?
Hi @Libra. Thank you for bringing this to our attention. Synthesized voices are definitely not welcome in the dataset. If you come across one, please vote no on the recording. I’ll speak to the team about some approaches we can try for preventing contributions of synthesized voices.
If you have some ideas for how to address it, we always welcome suggestions, either here or especially through opening an issue on the Common Voice GitHub. You can find issue templates here.
Hi, thanks for your reaction. Unfortunately, I don’t know how to find recordings like that automatically, because I’m not a programmer. The only one idea I have to exclude all recordings from the user-ids, who created these reported recordings, if they are made from the same accounts.
Also it looks it is an old problem of Russian dataset:
Synthesized voices are definitely not welcome in the dataset.
Copyright issues or there are any other reasons behind this decision? As they said in this answer by @don-Rumata:
RU: А, я думал, что это фишка такая. Типа больше вариантов голосов, в том числе и синтезированных. Ну чтобы Алиса и Сири друг друга лучше понимали.
EN: Oh, I thought it was a thing. Like more voice options, including synthesized ones. Well, so that Alice and Siri understand each other better.
that sounds like absolutely logic and relevant idea
MCV is collecting data on “how people speak (ideolect)”, not how machines speak. Including syntetic voices will downgrade the quality of a dataset. IMO:
- it is not ethical (it is based on someone else’s voice)
- it is degenerative for the models
- Of course there might be licensing issues (you cannot CC0 that you do not own)
- these systems generate a limited number of voices, adding them will bias the models
About the “horses and birds” as BG effects from games - these should be added as “augmentation” during training. More info here:
And about machines speaking to each other: IMO this is a very bad idea and will result in end of humanity.
If anybody would like to create a model based on voices generated by TTS, they can do it on their own, you would not need MCV for this.
Hi @Libra, yeah I would mostly echo what @bozden said. People are always welcome to download the CV dataset and use those methods for data augmentation on their end. There are certain research concerns that arise when using speech synthesis (TTS) for data augmentation, but balancing the impact of such decisions is left to those using the dataset. Using artificial methods for creating more data is best done from a dataset that doesn’t already contain artificially generated data.