What if people are using text-to-speech to record?

[DaDiRa]

I’ve come across recorded sentences with text to speech. Should I vote them positively or not?

[nukeador (Ruben Martin ❌ [away till Aug 31st])]

Hi @DaDiRa

Welcome to the community discourse!

This is interesting, I haven’t found this situation, did you have the chance to document which sentences were using this?

I would say that this is not ideal, since this is the same voice over an over again, so having more than 15 minutes of this voice is not super helpful. We really need at least 1000 different and diverse voices for each language, and definitely this is not very diverse.

[DaDiRa]

No I didn’t document it but I’ll do it from now on. This happened about 3 times so far in the 55 sentences I voted for.

[lissyx]

This is interesting, I haven’t found this situation, did you have the chance to document which sentences where using this?

I’m pretty sure this is something we already discussed about with @kdavis and the answer was a clear no as much as I can recall. Not only it’s going to not be very good for the dataset, but chances are that this is against the terms of use of the Text-to-Speech service.

[kdavis]

If the voice is indeed synthetic the clip should be marked as invalid, and I agree with @nukeador that…

Not only it’s going to not be very good for the dataset, but chances are that this is against the terms of use of the Text-to-Speech service.

[Michael_Maggs]

I’ve added this to the draft reviewing guidelines, here:

Discussion of new guidelines for recording validation

DRAFT GUIDELINES FOR REVIEWING RECORDINGS [Edited to include comments up to 13 March 2019] Problems with the written text Ignore minor problems of punctuation if they don’t affect the recording, but reject punctuation errors that may prevent a machine learning algorithm from properly recognising words: [Written text says:] “the giant dinosaurs of the Triassic, [The punctuation errors don’t affect the recording] [Written text says:] The giant dino”saurs of the Trias…

[Codigo_Logo_Programacao_e_Inteligencia_Artificial (Pedro Lima)]

Hi, so people are recording TTS clips, I’m rejecting them since it doesn’t make sense to have them in the dataset. I’m worried this will slow down the validation process of actual clips.

[nukeador (Ruben Martin ❌ [away till Aug 31st])]

@Codigo_Logo_Programacao_e_Inteligencia_Artificial how many of these have you found?

[Codigo_Logo_Programacao_e_Inteligencia_Artificial (Pedro Lima)]

@nukeador About 10 in a set of 150 clips.

[nukeador (Ruben Martin ❌ [away till Aug 31st])]

@gregor is there a way we can help people identify and flag these ones so we can identify who is sending these?

[gregor]

Unfortunately we don’t have flagging functionality yet, though it’s been requested a couple of times already.