The initial launch should download the default model.
If there's a problem with the initial download, you can download the needed files from: xTTS-V2 (you need to download: config.json, model.pth, speakers_xtts.pth, vocab.json).
This is how your xtts_model folder should look like :
So I run the server (via the exe) and nothing happens. Nothing is downloaded. I get asked: This is the host address the server will bind to Host to Bind [localhost]:
Some guidance for the less capable would be hand please! I can't get this to do anything.
I don't get any response in CHIM Discord, maybe you can help. I installed Mantella XTTC. And I just can't hear 90% of the speech. At the same time, dialogues are displayed in the console, these files also appear and play in the soundcache. But I can't hear them in the game. And some phrases are still audible. I've already racked my brain. Help!
Hey. I had (or still have? it seems gone now) the exact same situation.
I use an 1.3 Mantella XTTS Server on my laptop (RTX 2070) and Chim on my gaming tower (RTX 4090). Ah yes I use the VR version with a home made modlist (Vortex)
I used it with Herika and Mantella fine. Now I tried Chim and.... same situation: It was as if I only hear 90% of the voice lines that are generated. I also asked for help at the discord and no one responded.
I meanwhile deinstalled and reinstalled Chim for the third time and since then, my test runs apeared as if I hear 100%. But these tests had not been in buisy places. Only me and 2 followers. But this time it apeared 100% relyable.
Hi guys I have an RTX 5080 that isn't supported by the distributed PyTorch version. I tried to force a PyTorch upgrade using pip3 inside the internal XTTS folder but then the server crashes on startup. When I try to launch it via command it acts like a new install (and will probably overwrite the updated PyTorch files). Any ideas on how to make it work? Thanks
Just got my new rig today, exact same card, exact same problem, womp womp...
Update: After digging it could in theory be possible to run XTTS CPU bound by enabling CUDA_LAUNCH_BLOCKING=1 but I would probably advise against it as it is intended for debugging purposes and may do more harm than good so unless NVIDA finally decides to add CUDA support for the latest version of PyTorch we are SOL so to speak.
Late answer but I managed to make it work by rebuilding the XTTS API server using the Github repo and updating Pytorch libraries. I then compiled it to an exe that launched and worked properly on my 5080. I don't know much about Python so that wasn't easy but it's doable.
Be sure to have version 3.11.7 of Python installed, other newer versions weren't working for me
Follow the manual installation steps from https://github.com/art-from-the-machine/xtts-api-server-mantella
Before installing dependencies, in the requirements.txt file change the line "tts==0.21.3" to "coqui-tts[languages]==0.24.3"
I don't know why but I also had to add two other dependencies: "uuid" and "spacy==3.7.2"
Always in the same file, remove the lines "torch" and "torchaudio"
Once you've installed dependencies from requirements.txt, install PyTorch 2.7.0 using this command : "pip3 install torch torchaudio --index-url https://download.pytorch.org/whl/cu128"
Now you can either launch it directly from terminal or compile it to an exe using pyinstaller.
Hi thanks for the instructions, I started the api server in the terminal, but mantella crashes with error File "src\utils.py", line 15, in wrapper File "src\http\routes\mantella_route.py", line 56, in _setup_route File "src\utils.py", line 15, in wrapper File "src\tts\xtts.py", line 51, in __init__ File "src\utils.py", line 15, in wrapper File "src\tts\xtts.py", line 122, in _get_available_speakers AttributeError: 'list' object has no attribute 'get'
@leosdt I've tried to follow all these instructions, and doesn't work because the manual installation doesn't find the latent_speaker_folder. When tested with gradio it doesn't give the access to select language and model, that because the manual installation differs from the automatic, and doesn't give a base config.ini, and if I copy that from the automatic one, doesn't work anyway.
For those interested, I uploaded source code and binaries here : https://github.com/shmitix/xtts-api-server-mantella-blackwell
Can't really troubleshoot you guys as I haven't met these errors. I know at some point I had issues with relative and absolute paths so maybe you should look into it.
Is there any way to have Arabic Text-To-speach? I understand that xtts supports other languages, but how do I add Arabic? What other TTS tools can I use that support more languages?
Hi, great stuff, thanks! However I cant download the "speakers" from your file list, they are no longer active but archived. Only the latent is available. Without the wave files the whole thing does not work now, does it? Where can I download them? thanks for your help
After updating Mantella, I get this error.The NPC speaks with text, but no sound plays. This didn't happen with the previous version of Mantella. What's the cause?
Does anyone know if there is a video that explains how to create the custom voice model for an NPC? I followed the written instructions with the WAV audio but I can't get the JSON file to generate.
112 comments
If there's a problem with the initial download, you can download the needed files from: xTTS-V2 (you need to download: config.json, model.pth, speakers_xtts.pth, vocab.json).
This is how your xtts_model folder should look like :
└── xtts_models/ |
└── v2.0.2/
├── config.json
├── model.pth
├── speakers_xtts.pth
└── vocab.json
This is the host address the server will bind to
Host to Bind [localhost]:
Some guidance for the less capable would be hand please! I can't get this to do anything.
I use an 1.3 Mantella XTTS Server on my laptop (RTX 2070) and Chim on my gaming tower (RTX 4090).
Ah yes I use the VR version with a home made modlist (Vortex)
I used it with Herika and Mantella fine.
Now I tried Chim and.... same situation: It was as if I only hear 90% of the voice lines that are generated. I also asked for help at the discord and no one responded.
I meanwhile deinstalled and reinstalled Chim for the third time and since then, my test runs apeared as if I hear 100%. But these tests had not been in buisy places. Only me and 2 followers. But this time it apeared 100% relyable.
I need to do more tests in rooms with more NPC.
I have an RTX 5080 that isn't supported by the distributed PyTorch version. I tried to force a PyTorch upgrade using pip3 inside the internal XTTS folder but then the server crashes on startup. When I try to launch it via command it acts like a new install (and will probably overwrite the updated PyTorch files).
Any ideas on how to make it work?
Thanks
Update: After digging it could in theory be possible to run XTTS CPU bound by enabling CUDA_LAUNCH_BLOCKING=1 but I would probably advise against it as it is intended for debugging purposes and may do more harm than good so unless NVIDA finally decides to add CUDA support for the latest version of PyTorch we are SOL so to speak.
Now you can either launch it directly from terminal or compile it to an exe using pyinstaller.
thanks for the instructions, I started the api server in the terminal, but mantella crashes with error
File "src\utils.py", line 15, in wrapper
File "src\http\routes\mantella_route.py", line 56, in _setup_route
File "src\utils.py", line 15, in wrapper
File "src\tts\xtts.py", line 51, in __init__
File "src\utils.py", line 15, in wrapper
File "src\tts\xtts.py", line 122, in _get_available_speakers
AttributeError: 'list' object has no attribute 'get'
Have you set up anything else?
I've tried to follow all these instructions, and doesn't work because the manual installation doesn't find the latent_speaker_folder. When tested with gradio it doesn't give the access to select language and model, that because the manual installation differs from the automatic, and doesn't give a base config.ini, and if I copy that from the automatic one, doesn't work anyway.
Can't really troubleshoot you guys as I haven't met these errors. I know at some point I had issues with relative and absolute paths so maybe you should look into it.
TTS: Connecting to XTTS....
TTS: Could not connect to XTTS. Attemping to run headles server...
And it just stays like that. Any tips?
However I cant download the "speakers" from your file list, they are no longer active but archived. Only the latent is available.
Without the wave files the whole thing does not work now, does it? Where can I download them?
thanks for your help
thread_local object has no attribute data