Python scripts I made to make NNSVS labeling easier.
You will need Python for all of these to work, of course. I wrote this in Python 3.8.10
Here's a table for jumping for convenience
| lab2ust |
| database_stats |
| lab2audacity.py |
| check_labels.py |
| world_shift.py |
| harvest_frq.py |
| threaded_noise_remove.py |
The whole lab2ust folder will be put in the UTAU plugins folder.
Just follow the steps that the plugin gives you. It supports stringing together phonemes to notes. It supports the Japanese phoneme set that NNSVS, CeVIO, Sinsy, NEUTRINO, and Synthesizer V uses.
Japanese is coded in differently to support suffixes on vowels. This was mostly an idea that my friends and I thought of. In that case, Japanese hopefully supports VS where V is the vowel and S is the suffix. (e.g. falsetto could be aF, iF, and so on)
Other languages are supported by the languages.json file. Some languages that have public repositories have been added to this file. Adding a language would only require adding another definition of the language. I hope following the JSON format is easy as I didn't want to use YAML since that requires installing a library. Consonant splitting is still primitive as it only finds the center and splits that way.
It will always separate the phonemes pau, br, and sil.
This will only directly translate timing according to the BPM of the UST. It puts all notes at middle C (C4) when a .frq file is not provided.
Update 09/28/2021: lab2ust now quantizes notes. You need to specify the note length of the quantization though. Here's a list of note lengths for quantization. If you don't want quantization, just put 1.
| Quantize | Note Length |
|---|---|
| 128th note | 15 |
| 64th triplet | 20 |
| 64th note | 30 |
| 32nd triplet | 40 |
| 32nd note | 60 |
| 16th triplet | 80 |
| 16th note | 120 |
| 8th triplet | 160 |
| 8th note | 240 |
| Quarter triplet | 320 |
| Quarter note | 480 |
Update 03/02/2022: lab2ust can now read .frqs to automatically place the pitches of the notes. It is recommended to generate .frqs with moresampler for highest accuracy. PS: It can ONLY read .frqs so you would need to convert the .mrq to .frq. This can be done with frq editor. If you are OK with checking the frequencies instead, you can generate with speedwagon instead.
Update 08/02/2022: lab2ust will now check if there are parts of the label that only has consonants and is surrounded by the standalone phonemes pau, br, and sil. It will fuse them into one note to avoid zero length notes. lab2ust also supports custom language by allowing the user to input the vowels of the language comma-separated. This also accepts it if each phoneme is surrounded by either ' or ".
Update 08/10/2022: Improved language support by removing (almost all) the hardcoded languages from the code. Japanese is the only language that stays hardcoded because people have made resources specific to how I have developed the Japanese support. Basically, I made adding languages easier but kept Japanese the same cuz backwards compatibility.
This script is for getting some statistics for your NNSVS/ENUNU database. It reads all the USTs and LABs of the database (it will only count the LABs that already have USTs by default). You can drag and drop the database folder over the script like with lab2audacity, but this script also takes in additional arguments for settings. Here's the whole help documentation of the script, but this may be invoked with database_stats.py -h as well.
usage: database_stats.py [-h] [--all-labs] [--include-pau] [--skip-diphone] [--write-diphone] db
Calculates some statistics for your existing NNSVS database
positional arguments:
db The database's folder location
optional arguments:
-h, --help show this help message and exit
--all-labs, -l Include all LABs in the LAB pass.
--include-pau, -p Include pau phoneme in the phoneme tally passes.
--skip-diphone, -s Skip diphone density in calculations.
--write-diphone, -w Include diphone density in the .csv version.
This script will then save a stats.txt file inside the database folder you have passed to it.
Update 09/27/2022: This script will now generate a stats.csv file as well to import statistics into a spreadsheet format automatically.
This script can convert between Audacity labels (in .txt filetype) and HTS mono labels (in .lab filetype). Drag and drop the file over the script file and it will do the conversion. It cannot batch convert labels.
This script automatically connects the ends of each label to the start of the next label. If the label ends with -, it will exclude it in the HTS label.
This script does not do anything about timing with HTS mono labels.
It's exactly what the file says. It checks for wrong phonemes in every .lab file that it finds. It also checks if you have a label that has the same start and end timing. To run, put it in the folder where you keep your .lab files and just, run normally!
It checks Japanese labels by default, but if you want to change what language it checks, just open it up and change the phones list to the phoneme list you want it to be.
Update 06/15/2022: Added a feature that counts how many instances of each phoneme is present in the labels. The program will alert you if there is no instance of a certain phoneme by default but you may change it to be a certain threshold. At line 28 in the code (looks like if v == 0:) you may change it to something like if v <= n: with n being a threshold that you think is appropriate enough.
This script is for data augmentation (in a sense). It shifts all the .wav files in the folder that the script is in by the amount of semitones listed in the pitches list of the script. By default it is ±5 and ±12. You can change it by opening the script in a text editor and changing it yourself.
Example for changing the pitches:
#...
import os
import logging
pitches = [-12, -5, 5, 12] # This is the part to change.
# Let's say you want only ±3. You change it to pitches = [-3, 3]
def shift_pitch(path):
loc, _ = os.path.splitext(path)
directory, fname = os.path.split(loc)
# ...This script requires numpy, soundfile, and pyworld to be installed in your Python environment. You can just run pip install numpy soundfile pyworld in the terminal to install it.
UPDATE 06/14/2023: Switched to soundfile for dealing with .wav files.
This script generates .frq files using the Harvest F0 estimation algorithm from WORLD. It uses Python's multiprocessing module to speed up the process of generating .frq files. This script is definitely not restricted to usage for NNSVS only, as you can use it to make .frq files for your UTAU voicebanks as well. You may also drag and drop the folder over the script file to run it, or use the terminal to run it.
usage: harvest_frq.exe [-h] [--num-threads NUM_THREADS] path
Generate .frq files using WORLD's Harvest F0 estimation algorithm.
positional arguments:
path The path to a .wav file or a directory with .wav files.
optional arguments:
-h, --help show this help message and exit
--num-threads NUM_THREADS, -n NUM_THREADS
How many threads to use. Default is running single threaded. Input zero to use all available
threads.
This script also requires numpy, pyworld and soundfile.
WARNING: It seems to be generally safer to run harvest_frq without threading, as Harvest itself is quite CPU-intensive. If you think you can run it with multiple threads you may do it, but remember that you are willingly putting your computer at risk by doing so.
UPDATE 12/13/2022: Made this script support making a .frq file for a single .wav file through the same drag and drop behavior.
UPDATE 06/14/2023: Added an optional argument to limit the number of threads used and switched to soundfile for reading .wav files.
UPDATE 06/21/2023: Default to running without threading. Read warning as to why.
This script uses the Log-MMSE algorithm to denoise all .wav files in a folder, including its subfolders. It assumes that the first 120ms of each sample is noise. Performance is also dependent on the noise type as this is still an algorithmic noise remover.
usage: threaded_noise_remove.exe [-h] [--single-thread] [--num-threads NUM_THREADS] path
Denoises all wave files in a directory using the Log-MMSE algorithm.
Assumes the first 120 ms of the samples are pure noise.
positional arguments:
path The path to a .wav file or a directory with .wav files.
optional arguments:
-h, --help show this help message and exit
--single-thread, -s Run single threaded
--num-threads NUM_THREADS, -n NUM_THREADS
How many threads to use. Default is your thread count.
This script requires numpy, scipy and soundfile. The Log-MMSE implementation is a direct translation of this MATLAB code into Python.
Slight WARNING: threaded_noise_remove seems to be quite safe even with threading, but please do check if your CPU can handle it.
UPDATE 12/13/2022: Made this script support denoising a single .wav file through the same drag and drop behavior.
UPDATE 06/14/2023: Added an optional argument to limit the number of threads used and switched to soundfile for dealing with .wav files.