Compiled Windows guide with all issues fixed

[nitinmukesh]

https://youtu.be/hKb8lNuxIRU

Credits to @lethanner for his help in solving issues

[lethanner]

Are you really got perfect English pronunciation using Japanese voice models?
I can't understand any English words in my generated audios.

[nitinmukesh]

except for few words. I haven't done enough testing though, will try this weekend.
I included 2 songs in the video.

[lethanner]

Japanese voice models

Pardon. Japanese voice are not really used in the inference. It is only used during initialization.
See #17 (comment) for explanation.

[nitinmukesh]

I read your detailed analysis, thanks for sharing it.

[Sxrjs2010]

Thanks

[sujitvasanth]

thanks for this... I also created a summary of what I needed to make it work on windows for someone on the discord channel that was struggling - much. much easier to run on ubuntu..(!)

To create a windows local server us.txt

To create a windows local server using local GPU of https://huggingface.co/spaces/ASLP-lab/DiffRhythm

1. create a python 3.12.9 venv
   ==============================
   download and install python 3.12.9 in default directory with add pip and add pythonpath options
   it can be found at python https://www.python.org/downloads/windows/
   open powershell
   py -3.12 -m venv diff-gradio
   python -m venv diff-gradio
   Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser
   diff-gradio\Scripts\activate

2. clone the hugging face repo
   ==============================
   download and install git from https://git-scm.com/downloads/win
   git lfs install
   git clone https://huggingface.co/spaces/ASLP-lab/DiffRhythm
   cd C:\Users\44741\DiffRhythm <--- need to change to your local path
   dont install requirements.txt yet

3. install libraries for windows
   ================================
   a) triton
   python --version
   download the right version of triton.whl from https://huggingface.co/madbuda/triton-windows-builds/tree/main
   pip install C:\Users\44741\DiffRhythm\triton-3.0.0-cp312-cp312-win_amd64.whl <--- need to change to your local path
   b) pytorch with cuda
   pip uninstall torch
   pip cache purge
   nvidia-smi (to know your version of cuda)
   go to https://pytorch.org/get-started/locally/ and get your pip install (but change pip3 in command to pip)
   e.g. pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
   c) requirements.txt
   cd C:\Users\44741\DiffRhythm <--- need to change to your local path
   pip install -r requirements.txt
   d) cuda version of onnyxruntime
   pip install onnxruntime-gpu

4. make the file alterations to the cloned repo so it works for windows
   =======================================================================
   Filename Folder
   init.py diffrhythm\g2p\g2p
   g2p.py diffrhythm\g2p\utils
   g2p_generation.py diffrhythm\g2p
   infer_utils.py diffrhythm\infer
   app.py root

diffrhythm\g2p\g2p_init_.py, line 28
with open(vacab_path, "r", encoding="utf-8") as f:

diffrhythm\g2p\utils\g2p.py, line 63
with open("./diffrhythm/g2p/utils/mls_en.json", "r", encoding="utf-8") as f:

diffrhythm\g2p\g2p_generation.py, line 117
with open("./diffrhythm/g2p/g2p/vocab.json", "r", encoding="utf-8") as f:

diffrhythm\infer\infer_utils.py, line 105
with open('./diffrhythm/g2p/g2p/vocab.json', 'r', encoding='utf-8') as file:

app.py, comment out lines 12 and 33, add environmental variables after line 11
12 #import spaces
33 #@spaces.GPU(duration=20)
12 os.environ["PHONEMIZER_ESPEAK_PATH"] = r"C:\Program Files\eSpeak NG"
13 os.environ["PHONEMIZER_ESPEAK_LIBRARY"] = r"C:\Program Files\eSpeak NG\libespeak-ng.dll"

5. install espeak-NG
   ====================
   see https://www.youtube.com/watch?v=J8FejpiGcAU

a)Download and install eSpeak NG
https://github.com/espeak-ng/espeak-ng/releases

b) Install MbrolaTools35
https://archive.org/details/MbrolaTools35

c) Download additional file
https://github.com/thiekus/MBROLA/releases/download/3.3/mbrola_build_3.3_rev2.zip
Copy all files from mbrola_build_3.3_rev2\Win64\Release
Paste inside C:\Program Files\eSpeak NG

d) Download MBROLA-voices
https://github.com/numediart/MBROLA-voices
Download jp1, jp2, jp3
C:\Program Files\eSpeak NG\espeak-ng-data\mbrola\

e) Download MBROLA-voices
https://github.com/user-attachments/files/19094349/mbrola_ph.zip
C:\Program Files\eSpeak NG\espeak-ng-data

6. run the gradio server
   ========================
   in the venv
   cd C:\Users\44741\DiffRhythm <--- need to change to your local path
   python app.py

[NZqian]

Japanese is now temporarily removed, so I believe this issue is solved XD

[erhan-]

Doing just a fresh install on Windows with current master and it is not solved:

RuntimeError: failed to load voice "ja"

Besides this is more supposed to be a guide and should be in a markdown file instead of an issue.

Open .venv/Lib/site-packages/phonemizer/backend/espeak/wrapper.py and replace line 240 with
if (voice.language not in available) or (voice.language == 'ja'):

[nitinmukesh]

This was closed for a reason. With new updates released yesterday some issues are fixed and no longer require manually downloading of files.

[erhan-]

@nitinmukesh I posted this for a reason. The error appears still with your commits.

See your commits on the left.

With my proposed edit it works at least as a workaround:

[erhan-]

I created a pull request here to fix it: #44

[nitinmukesh]

@erhan-

I have not added any commits in this repo. It seems you are using this repo, whereas I created a different repo for Windows and that is DiffRhythm-windows.

I can see in your code it is DiffRhythm and not DiffRhythm-windows.

[erhan-]

Sorry for the mixup. I though the response came from NZqian. The content is still valid though and the PR is required.
Not sure what you mean with DiffRhytm-windows.