[Generative AI] VoiceGPT

[AiKodex]

@Ictcmn

Also, will there be a way to batch-process lines of text? 500 isn’t loads of characters at once, but maybe being able to batch process one after the other would help with this limitation

You can go over 500 characters in the offline mode, but the voice sometimes becomes unstable after 750 characters and you have to discard the entire sample due to that. We have had some 2000 character extracts that were processed to clips successfully too. We are still actively developing this tech.

[AiKodex]

@Gamingbir

Its mentioned that I can make accents. How do I make use of accents?

Hey Gamingbir,

You can create different accent combinations by typing the text in the main language you want the voice to be in and selecting a different language to create an accent onto the voice. Please refer to the first page of the documentation for some samples. This is not available in the offline mode currently.

[AiKodex]

@BenoitFreslon

Hello there, I got the same issue when I try to generate a voice using local model.

PythonException: only integer tensors of a single element can be converted to an index
File "Library/PythonInstall/lib/site-packages/vgpt/tts.py", line 256, in inference
ref = alpha * ref +(1 - alpha)  * ref_s[:, :128]
File "<string>", line 11, in <module>
File "<string>", line 1, in <module>

Hello Benoit,

There is a high chance that this is due to the usage of a EU based decimal comma than a decimal point. @Nightkin99 in his earlier post was able to solve this issue by changing the his system settings to decimal point.

@Nightkin99

but the real culprit was indeed the decimal separator! My OS is Windows 10 and it is set to use a comma as the decimal separator, for testing reasons. In my game, when I create a save, I need to be sure that all the numbers are stored with a dot, so I use “InvariantCulture” when calling ToString() for saving and TryParse() for loading. Using a comma on my system helps me spot parsing errors early, instead of waiting for a European player to report the bug because they don’t use the “standard” decimal separator (a dot).

So I changed my system to use a dot instead of a comma, and it worked! I don’t want to do that, though. My system should keep using commas (whether I like it or not).

[TriCounter]

How does the option for DeepVoice character extensio work in practice?

“Also can be used to extend the character count of DeepVoice up to 560,000 char/mon”

[BenoitFreslon]

@TriCounter

How does the option for DeepVoice character extensio work in practice?

@AiKodex

Hello Benoit,

There is a high chance that this is due to the usage of a EU based decimal comma than a decimal point. @Nightkin99 in his earlier post was able to solve this issue by changing the his system settings to decimal point.

“Also can be used to extend the character count of DeepVoice up to 560,000 char/mon”

Thanks for the fix. However, if players use european decimal separator, it could be a problem?

[AiKodex]

@BenoitFreslon

Thanks for the fix. However, if players use european decimal separator, it could be a problem?

The fix was a temporary one. We will be releasing an update which hopefully fixes this issue.

[AiKodex]

@TriCounter

How does the option for DeepVoice character extensio work in practice?

“Also can be used to extend the character count of DeepVoice up to 560,000 char/mon”

Hello TriCounter,

You can simply load your DeepVoice sample into the VoiceGPT voice field and use the voices from there. However, there will be a few changes in the voice such as voice quality and sampling rate from 44kHz to 24kHz.

[AiKodex]

[Announcement]

We have updated VoiceGPT to v1.2.1.
Since the offline mode has received tons of attention, we have implemented the following changes:

• Increased the character count to 1000 characters for a single inference
• Solved the issues with EU decimal comma thanks to the suggestions by @Nightkin99
• Bug fixes with the text processing and the model parameter processing

[BenoitFreslon]

@AiKodex

[Announcement]

We have updated VoiceGPT to v1.2.1.
Since the offline mode has received tons of attention, we have implemented the following changes:

• Increased the character count to 1000 characters for a single inference
• Solved the issues with EU decimal comma thanks to the suggestions by @Nightkin99
• Bug fixes with the text processing and the model parameter processing

Nice. Waiting for the deployment on the Asset Store ^^.

[AiKodex]

@BenoitFreslon

Nice. Waiting for the deployment on the Asset Store ^^.

Hey Benoit,

The asset is live on the asset store now. We were able to also clean up the UI and remove some redundancies from the code as well in v1.2.2.

[Nightkin99]

Thanks for the update! I confirm that VoiceGPT now works with the decimal separator set to comma as well.

[crushed_unity_dreams]

thanks for adding offline!!! amazing!!!

[AiKodex]

[Announcement]

Automatic Quota Reset to 500,000 characters. [Quota changes]

500,000 characters allotted for the period 1-05-2024 to 31-05-2024.

[BenoitFreslon]

Hello,

Could you please provide an example script with the local model to generate AI voices on the fly in the build?

Thanks.

[AiKodex]

@BenoitFreslon

Hello,

Could you please provide an example script with the local model to generate AI voices on the fly in the build?

Thanks.

Hello BenoitFreslon,

Sure, we can send you the modified script by email. You can simply remove the AssetDatabase API / Editor API, make an interface / Scene GUI to handle local inferences and use that for inferencing the model.

[BV_G]

First off, amazing job with VoiceGPT! I’ve been having a blast using it. I do have a couple of questions and was hoping you could help me out.

I’m really interested in getting more languages to work in offline mode. Right now, it seems like only English is supported. Are there pre-trained models available for other languages that I can use offline? If so, how can I get my hands on them?

Also, I’m curious about training or finetuning the models for different languages. Is this possible with the current setup? If anyone’s done it, I’d love to hear how you did it or if there are any resources or tutorials you can point me to.

Any advice on the best practices for training or finetuning these models would be awesome too. What kind of datasets should I be looking for, and are there any specific requirements I should know about?

[AiKodex]

[Announcement]

Automatic Quota Reset to 500,000 characters. [Quota changes]

500,000 characters allotted for the period 01-06-2024 to 30-06-2024.

[Blinxer]

Hello! I ordered your product yesterday and it’s great - however, I need help with its runtime functionality. Do you have any example projects or scripts that show how to set VoiceGPT up such that it reads text on a Canvas? I can’t figure out the Example scene.

[AiKodex]

@Blinxer

Hello! I ordered your product yesterday and it’s great - however, I need help with its runtime functionality. Do you have any example projects or scripts that show how to set VoiceGPT up such that it reads text on a Canvas? I can’t figure out the Example scene.

Hello Blinxer,

For VoiceGPT it is not much different. We have sent you an email with detailed instructions on how you can use to generate audio in runtime and build.

Basically, you can take a closer look at the “Generate()” function which is the most important when it comes to generating the audio file. You then need to perform CRUD on the platform you want to build on. You can modify it to suit your needs whether you are building a game or automated voice over clips - the same as DeepVoice.

[AiKodex]

[Announcement]

Automatic Quota Reset to 500,000 characters. [Quota changes]

500,000 characters allotted for the period 1-07-2024 to 31-07-2024.