How to generate pitch_stats.npy file?

[RamakrishnaChaitanya]

I'm trying to finetune the model on openslr dataset. I followed all the steps mentioned in the documentation. However, I'm getting many issues. I started resolving them one by one. And i got stuck at this particular error -

self.mean = stats["mean"].astype(np.float32) in load_stats function of dataset.py file.

I went through the compute_statistics.py and executed the same, assuming that the output file is pitch_stats.npy. Is this assumption, correct? If yes, the fields that are being stored in the dictionary are only mel_min, mel_max, linear_min, linear_max. So, how to address the stats["mean"] issue?

stats = {} stats["mel_mean"] = mel_mean stats["mel_std"] = mel_scale stats["linear_mean"] = linear_mean stats["linear_std"] = linear_scale

Should we change the names of the keys? If yes, what keys should we change? Also, in addition to compute_statictics.py, should we execute any other files before training using the train_tts.py file?

[eginhard]

Could you share the following details:

• model you're fine-tuning
• command you're running
• full error message
• are you using the fork (pip install coqui-tts instead of pip install TTS)?

Where does it say that you should run TTS/bin/compute_statistics.py? You shouldn't need it and the correct statistics should be computed automatically as part of the training process.

[RamakrishnaChaitanya]

Hi, I'm finetuning the FastPitch model. The command that I'm executing is python train_tts.py --config_path IndicTTS\en+hi\fastpitch\config_2.json --restore_path Best_models_IndicTTS\en+hi\fastpitch\best_model.pth on the console.

The error message I'm getting is shown in the attached picture.

And I'm using pip install TTS. I just explored the ways to create pitch_stats.npy file and found that compute_statistics output a .npy file. I assumed that the output generated using compute_statictics.py is pitch_stats.npy.

If i don't execute compute_statictics.py before hand to generate a .npy file, then I was getting the following error

[eginhard]

It's better to use the fork, i.e. do pip install coqui-tts, this repo is not maintained anymore.

I think the error occurs because the D:/.../f0_cache folder (you only shared an image, so I can't copy the path...) exists, but that folder doesn't contain pitch_stats.npy. Currently the statistics are only auto-computed when the whole folder doesn't exist. So it should probably just work after deleting the f0_cache folder.

[RamakrishnaChaitanya]

Thanks @eginhard for your feedback. I was able to start the training after deleting the existing f0_cache folder. However, i observed that multiple *.npy files are getting created in the f0_cache folder in the initial stage of training. I'm trying to finetune the fastpitch model on a dataset which consist of 52k sentences and i realized that .npy files are getting created for each of the sentence. I later observed that this process of generating *.npy files in the f0_cache is dead slow even though i'm using a GPU (48GB). Is this expected or are there any better ways to generate the .npy files?

[eginhard]

Yes, it creates one .npy file with the pitch data for each audio file. It will take some time and can't be sped up by GPU, but only needs to be done once in the beginning. You can set a higher value for precompute_num_workers to run it on multiple CPU cores.