MD's Skyrim Voice Tools

[MellowDrama]

12 hours ago, Courier07 said:

Any chance I could get some help or am I figuring this out myself? I understand nearly everything else, the only thing I don't get is the jobs field. It points to some example txt file, and so what am I supposed to do with that?

Hi sorry keep forgetting to actually answer you.

The installer generates a .pas file

It's just a script that SeEdit can run with äpply script" on an esp or dialogue topic(s)

This should automatically open the actual thing.

And you can click on generate all lines.

If not
you can change the source file to another file as long as you keep the same format, you will generate the lines from your own .txt

 

Hope this helps

[MellowDrama]

16 hours ago, StuffedNinga said:

It works great but it seems to want to convert each generated wav file to fuz immediately, which seems to slow down the generation process.  Each audio file takes about 30 seconds to complete a fuz file whereas other TTS generators take only about 4 seconds for me to complete a wav file.  Would it be possible to only generate wav files and leave the fuz generation process for the end, as a separate process, after all wav files have finished generating?  This would also help folks listen to individual audio files for issues and easy editing with Audacity before finalizing them as a fuz.

 

You also might not want to make it too easy to generate audio to completion as this completely bypasses quality control.  Mass generated audio will sound mostly good but you'd still need a human to listen to the generated files for errors and proper emotional context before finalizing it.  A one click production, like you currently have, is a breeze to run, but it also means folks are less likely to check the generated audio to make sure it's actually good.

Not sure why i wasnt already doing that. Good idea.

Also if you do two or more tries it will generate multiples of the same line and let you chose which one to keep.

Edited July 31 by MellowDrama

[Courier07]

21 minutes ago, MellowDrama said:

Hi sorry keep forgetting to actually answer you.

The installer generates a .pas file

It's just a script that SeEdit can run with äpply script" on an esp or dialogue topic(s)

This should automatically open the actual thing.

And you can click on generate all lines.

If not
you can change the source file to another file as long as you keep the same format, you will generate the lines from your own .txt

 

Hope this helps

Oh yeah I figured that out after reading through your posts one more time. You actually had it clearly stated in plain text, I'd missed it. Now to generate the voice, I need 30 secs of that new voice stitched together, and that type added to the type file. Is that it? Also where do I put the sample voice line, that is, the lines generated for another voice which I want to use as the example? clone from?

[MellowDrama]

10 minutes ago, Courier07 said:

Oh yeah I figured that out after reading through your posts one more time. You actually had it clearly stated in plain text, I'd missed it. Now to generate the voice, I need 30 secs of that new voice stitched together, and that type added to the type file. Is that it? Also where do I put the sample voice line, that is, the lines generated for another voice which I want to use as the example? clone from?

I added this button for you

You can keep transcript empty but it works better with a transcript (actually that is turned off but it should work in the next update)

[StuffedNinga]

Most dialogue has an emotional tag to them which are Anger, Disgust, Fear, Happy, Neutral, Puzzled, Sad, Surprised.  Would it be possible for your program to auto detect these tags and load a reference wav file with a similar emotional context to them?  Like if a dialogue is labeled as Angry, the program will look for a reference wav that is also labeled as Angry to generate it's dialogue from?

 

Dialogue also has an emotional intensity level from 0 - 100 which could help further define generated audio.

[MellowDrama]

3 hours ago, StuffedNinga said:

Most dialogue has an emotional tag to them which are Anger, Disgust, Fear, Happy, Neutral, Puzzled, Sad, Surprised.  Would it be possible for your program to auto detect these tags and load a reference wav file with a similar emotional context to them?  Like if a dialogue is labeled as Angry, the program will look for a reference wav that is also labeled as Angry to generate it's dialogue from?

 

Dialogue also has an emotional intensity level from 0 - 100 which could help further define generated audio.

Nah Qwen doesn't support that maybe if they add support for it one day but currently the only way to put that in would be extremely clunky

 

[MellowDrama]

On 7/30/2026 at 10:58 PM, StuffedNinga said:

It works great but it seems to want to convert each generated wav file to fuz immediately, which seems to slow down the generation process.  Each audio file takes about 30 seconds to complete a fuz file whereas other TTS generators take only about 4 seconds for me to complete a wav file.  Would it be possible to only generate wav files and leave the fuz generation process for the end, as a separate process, after all wav files have finished generating?  This would also help folks listen to individual audio files for issues and easy editing with Audacity before finalizing them as a fuz.

 

You also might not want to make it too easy to generate audio to completion as this completely bypasses quality control.  Mass generated audio will sound mostly good but you'd still need a human to listen to the generated files for errors and proper emotional context before finalizing it.  A one click production, like you currently have, is a breeze to run, but it also means folks are less likely to check the generated audio to make sure it's actually good.

It's 4 times faster now on my CPU (not counting the actual converting but that doesn't take long) Any more ideas?

 

[MellowDrama]

11 hours ago, Xarcellmo said:

I got the same problem with sox.

You dont need sox

its something from qwen but i used something else it's fine.

[StuffedNinga]

Just now, MellowDrama said:

You dont need sox

its something from qwen but i used something else it's fine.

If you didn't already do so with your latest update, the conversion to fuz should be a manual button you press at the very end.  The idea is to have the wav files available for preview and adjustment before being finalized as a fuz.  This is also the only way to catch typos, is if a human listens to the audio.

 

Perhaps before the program starts, the generated sse csv file can be previewed?  Even notepad has a basic spell checker now though it doesn't have a grammar checker.  But going through the file before generating to fix errors would save a lot of trouble prior to generating audio.

 

If the ssedit script doesn't already do so, you might want to organize audio by voice type.  This might help the program cut down on constantly switching between reference audio when batch generating and might come with a speed bump.

[StuffedNinga]

1 hour ago, MellowDrama said:

Nah Qwen doesn't support that maybe if they add support for it one day but currently the only way to put that in would be extremely clunky

 

Quen uses a reference audio to do it's thing.  What you would need is a reference audio of each type of emotion so like FemaleNord Happy, FemaleNord Sad.  You'd put those in a separate folder.  Then when you generate for csv in ssedit, you'd also separate out these emotions so a file for happy and a file for sad.  When you load the happy file, you'd transfer the happy audio to the folder of reference audio and when you load the sad file you'd transfer over the sad audio.  It's a bit convoluted but I do this manually myself and the results are much better than using a single generic reference audio.

[MellowDrama]

35 minutes ago, StuffedNinga said:

Quen uses a reference audio to do it's thing.  What you would need is a reference audio of each type of emotion so like FemaleNord Happy, FemaleNord Sad.  You'd put those in a separate folder.  Then when you generate for csv in ssedit, you'd also separate out these emotions so a file for happy and a file for sad.  When you load the happy file, you'd transfer the happy audio to the folder of reference audio and when you load the sad file you'd transfer over the sad audio.  It's a bit convoluted but I do this manually myself and the results are much better than using a single generic reference audio.

maybe just maybe

 

42 minutes ago, StuffedNinga said:

If the ssedit script doesn't already do so, you might want to organize audio by voice type.  This might help the program cut down on constantly switching between reference audio when batch generating and might come with a speed bump.

it does do that
 

44 minutes ago, StuffedNinga said:

If you didn't already do so with your latest update, the conversion to fuz should be a manual button you press at the very end.  The idea is to have the wav files available for preview and adjustment before being finalized as a fuz.  This is also the only way to catch typos, is if a human listens to the audio.

 

Perhaps before the program starts, the generated sse csv file can be previewed?  Even notepad has a basic spell checker now though it doesn't have a grammar checker.  But going through the file before generating to fix errors would save a lot of trouble prior to generating audio.

That is what the last update did

multiple tries already had it basically but now single try has it too

[Xarcellmo]

Alright, I got the software to work. However, maybe I'm missing something, but it seams to me that the generated audio gets named in a way that the game doesn't actually recognize. Like, I tested the output in the game and I couldn't hear anything. Infact, trying a script that generates a .csv file from the same .esp in order to get audio with xvasynth leads to outputs with different names that I can actually hear in game.

[MellowDrama]

25 minutes ago, Xarcellmo said:

Alright, I got the software to work. However, maybe I'm missing something, but it seams to me that the generated audio gets named in a way that the game doesn't actually recognize. Like, I tested the output in the game and I couldn't hear anything. Infact, trying a script that generates a .csv file from the same .esp in order to get audio with xvasynth leads to outputs with different names that I can actually hear in game.

can you tell me which names it used including the rest of the file location for both?

[Xarcellmo]

26 minutes ago, MellowDrama said:

can you tell me which names it used including the rest of the file location for both?

For example, the xvasynth script gives me "sound/Voice/TestDialogue.esp/FemaleArgonian/1testquest_1testquestdialo_00000D64_1.wav", which is also how the Creation Kit would name the audio, meanwhile with this tool I get "Sound/Voice/TestDialogue.esp/femaleargonian/1testquestdialoguetopic_02000d64_1.wav".

[necol]

11 hours ago, Xarcellmo said:

Alright, I got the software to work. However, maybe I'm missing something, but it seams to me that the generated audio gets named in a way that the game doesn't actually recognize. Like, I tested the output in the game and I couldn't hear anything. Infact, trying a script that generates a .csv file from the same .esp in order to get audio with xvasynth leads to outputs with different names that I can actually hear in game.

Unfortunately, I’m experiencing the same issue. The voice generation itself works perfectly, but the game does not recognize the generated voices (English version). Additionally, multilingual support is still not working, particularly for Russian.

[rsunmcoxuxypqwuvfr]

Playing around with it since yesterday, for now with generate-qwen-voice-fuz-batch.bat, I was pleasantly surprised by the gwen mimic ability to change the resulting voices as something close and similar from the source material (sounding like them).

 

For those, wanting to try, be aware, that (i tried with French Language):

• Hugging Face model (model.tensor) is a 3.8Go + 1.8Go models to install, it takes times and storage in c:\ python appdata subfolders to get and store
• I needed at least 10 to 13Go of available RAM (or pagefile.sys available) to run the voice inference, otherwise it may fails silenty/instantly (once i got a bluescreen, my system was out of memory out my 16Go)
• compared to XVASynth, the femaleYoungEager was not really sounding like Y'solda french dialog, but was decent on its own

I tried as well to create some custom voices (.mp3 files + helper transcripts files) from other games/video, and the cloning was getting something sounding a bit similar to the source.

 

So my questions are as follows:

• Any way to sped up the  inference/voice building on something like an Nvidia 1060? (i need to wait like 4-5min per line with a 50% CPU usage)
• From what i tested, only the transcript section in _transcript files are read, other fields are mostly the result of one your tool not shipped with the voice tools, time segments veracity or precision was not impactful at all on result/voice accentuations.
• Is there a minimum/essential coverage of sounds/words to cover in the 30sec of trimmed voice mp3 file to characterize a source voice? Should we exclude repeated words/sentences to not undo/average the cloning process of each word pronunciation?
• Does putting a mp3 file of something longer than 30sec (like more than 1min+) something negatively impactful for the resulting line or the generation time?

 Thank you for putting together the tool and scripting!

[MellowDrama]

16 minutes ago, rsunmcoxuxypqwuvfr said:

• Does putting a mp3 file of something longer than 30sec (like more than 1min+) something negatively impactful for the resulting line or the generation time?

 Thank you for putting together the tool and scripting!

I just got a 2070 super so i can start testing the CUDA stuff (I've a got r7900xt without cuda) but on the CPU it shouldnt take more than a minute per line.

There's no real minimum

[Xarcellmo]

I'd like to report that the generated filenames are still incorrect despite the update. This time I get "dialogue_1testquestdialo_00000d64_1" instead of "1testquest_1testquestdialo_00000D64_1".

[MellowDrama]

On 8/4/2026 at 1:08 PM, Xarcellmo said:

I'd like to report that the generated filenames are still incorrect despite the update. This time I get "dialogue_1testquestdialo_00000d64_1" instead of "1testquest_1testquestdialo_00000D64_1".

Can you test this? My skyrim is acting up.

[Xarcellmo]

I tested the lates version, the names are still incorrect.