Persian Piper TTS

[gyroing]

Hi
new Persian Piper model :
Persian Piper TTS based on persian-tts-dataset-male dataset
voice model training is under progress for finetune
@karim23657
Karim, thank you for the guidance on introducing the dataset, I think this is the best Persian model made.

[karim23657]

Thank you @gyroing ,For your amazing model, I added your model to list.
Could you please give me more information about finetuning process, which model was your base ?
Please share training code with us. 🤩

[gyroing]

I use google colab pro for training the model, I share the Jupyter Notebook ipynb file here, the main model is still in training progress then i will start finetuning, ,
12 years ago, I was involved in the Persian Espeak team with Shadyar Khodayari. I added many words there. In my opinion, currently, for every model that uses espeak phoneme, the vocabulary list should also be upgarded to the espeak fa_list and recompiled.
When (piper farid voice) this model reach to epoch200, I will share Persian model cheakpoin ".ckpt" links in huggingface.
Piper training for Colab

11 years ago i also convert persian espeak to javascript with emscripten :
Persian Espeak Chrom extension

@karim23657
Plaese see intersting job by jozefchutka:
piper-wasm
Piper Online Webassembly

it is possible to use any piper model locally in chrome extension with No conflicts with cors origin policy

[gyroing]

The model has been updated, a noticeable change can be seen
Now in epoch 300
till now has been consumed 25 hours for training with google (A100 40gig gpu) best gpu unit. around 25 hours.
Persian-Piper-Model-gyro

[karim23657]

@gyroing , I really appreciated for your great work on Persian espeak,

• i cant open your colab notebook , i think its because of sharing settings , i think it is better to upload it directly in your comment
• for piper-wasm , how can i help you?
• And i think the problem of most persian tts system is phonemizer, any colaboration or idea for this

[gyroing]

@gyroing , I really appreciated for your great work on Persian espeak,

* i cant open your colab notebook , i think its because of sharing settings , i think it is better to upload it directly in your comment

* for [piper-wasm](https://github.com/wide-video/piper-wasm) , how can i help you?

* And i think the problem of most persian tts system is phonemizer,  any colaboration or idea for this

Herewith i attach the folder link for related ipynb files:
ipynb for training and export

• the correct metadata.csv is also added in the folder . should be replace with persian-tts-dataset-male one after unzip
• google Colab t4 GPU is not enough for larger batch and will take hours, is better to upgrade Colab Pro and use A100 GPU
• after preparing dataset it will generate audio_cache (35gigabyte) the audo_cache folder will be erased with related session when you Close Google Colab page ,so you need to prepare it again. i suggest you to increase googledrive size and change the path of audiocahe folder from Colab code to GoogleDrive for future time saving (around 30 min )
• the checkpoint size will be 807 MB

for piper webassembly i manged the code to run locally in any modern browser i will send the necessary files.

[gyroing]

WebAssembly Sherpa onnx TTS with Persian Piper Model :
https://github.com/gyroing/Persian-Piper-TTS-WebAssembly

[karim23657]

@gyroing , Could you please share your checkpoints for finetuning new voices?