This is a CLI app that wraps the CLI app for TorToiSe TTS to be... more.

• Support for a voice.json file for voices so you can save your tunings for your custom voices and forget half the CLI settings.

• Process a single line, a text file as one paragraph, or a text file with each line being run separately pre-processed so TorToiSe can do work in nicer smaller batches - which helps with low VRAM hardware and avoiding some pitfalls where it will start stuttering, repeating itself, or skipping entire statements in longer texts.

• Writes a job.json file with each output so it can be re-ran as it was.

• Interception and rewriting of the progress output that did not work well on Windows, or if you ever resized your terminal window, or looked at it too hard on accident.

• Some pre-emptive error checking. Mostly prevention of shooting yourself in the foot with settings that end up crashing with a Python stack dump.

• Prevent you from just overwriting the same files over and over when you wanted to bulk process 100 voice lines, but with 3 samples each, such you would have woke up the next morning to only 3 files.

• If you use conda there is a reminder if you failed to remember to activate the conda environment. This can be disabled in tort.json.

• Dry-Run mode where it just shows you what is going to happen without doing any of it.

• A confirmation before actually doing it, showing you what it is going to do, which can be bypassed with a flag.

This is all done just with wrapping and checking things before commiting commands to TorToiSe. No way they'd been happy with the amount of reformatting I would have done with a pull request.

Tested on Windows 10 and Mint Linux 21. Both using conda hosting Python 3.9 environments to escape the Python minor version and package backward compatiblity breakage hell. I am certain an actual Python person would be able to navigate it without it though just fine.

• TorToiSe TTS (https://github.com/neonbjb/tortoise-tts)

• PHP 8.1

• Composer for PHP (if not using PHAR)

Grab the latest PHAR from Releases and extract it.

From within the tort directory install TorToiSe the way they instruct. You should spend time to get it to work there with the do_tts.py to prove it works or else none of this is going to work either. Do not skip over the bits about PyTorch, double so if you want GPU processing to actually work.

Here is the current path to success including dealing with how Python people are the best at constantly breaking their packages BC with minor point releases such that their package manager admits to being a meme:

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.

• cd tort
• git clone https://github.com/neonbjb/tortoise-tts
• cd tortoise-tts
  • conda create -n tort python=3.9.16
  • conda activate tort
  • pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
  • conda install -c conda-forge cudatoolkit pysoundfile
  • pip install -e .
  • pip install transformers==4.29.2
  • python setup.py install
  • cd ..

Next copy the voices folder out of tortoise-tts/tortoise/voices into the root of this project so you can edit and add configuration files to them. That way you can update TorToiSe at-will without messing up your progress.

• cp -r tortoise-tts/tortoise/voices .

The final directory structure should look like this:

• tort
  • tortoise-tts
  • voices
  • tort.phar
  • tort.json

If you are not using conda or do not want to be reminded when you failed to activate an environment, change the appropriate setting in tort.json.

Start by checking out this repo.

• git clone https://github.com/bobmagicii/tort-php-cli

Install the dependencies.

• cd tort-php-cli
• composer install

Then follow the exact same instructions as the Preferred Method skipping the first line.

This will dump its help command.

• php tort.phar

This will dump the help for the generation command.

• php tort.phar help gen

List all the voices that we were able to find. Probably important to check that first before going nuts.

php tort.phar voices

Run a single line of text on the default voice which I decided should be train_atkins.

php tort.phar gen "this is a test."

Run a single line of text with a custom voice.

php tort.phar gen "this is a test." --voice=bob

Run a text file as the input. This has TorToiSe handle the entire text file in the manner it decides to. This generally means it will break it into smaller chunks and run them in sequence.

php tort.phar gen --file=path\to\textfile.txt

Run a text file as the input, but have Tort pre-process it breaking it up by line and having each line run by clean invocations of TorToiSe. This helps a a lot with voices that start suttering, repeating, or skipping. It also helps with VRAM usage.

php tort.phar gen --file=path\to\textfile.txt --lines

Run with more Iterations. This actually defaults to four and if I am being honest I can rarely tell a difference between four and ninty-six with clean sound sources, other than a difference of like 10 minutes of life.

However each voice dataset is unique and often some do better with fewer or moreser you just gotta play around to get your stuff be good. I got one voice that sounds best at 62 iters, but then my Master Chief voice sounds better at 4 with him getting more and more southern the more iters added.

php tort.phar gen "this is a test." --iters=24

Run with a specific seed so you can repeat the same generations. I have observed that the same seed will not produce the same outputs on across different hardware. My NVIDIA 3060 and 980 both produced different outputs from the same seed.

php tort.phar gen "this is a test." --seed=69

Generate a new config file for your voice so you can forget all the tuning values that are listed by the tort help gen command later on. It will create a voice.json file in that voice folder with the tuning values that have the most effects and from that point on if you do not provide any CLI overrides, those setting will get used with that voice.

php tort.phar voiceconf <voice>

Tort includes a queue service that can be used to queue up many jobs and have them all processed in series. Jobs can be thrown at it from any angle and it will just grind them out in the order it got them.

Additionally, it will write the queue to disk so it can resume where it left off if something happens, or you need to stop it so you can have system resources back.

Run the queue server. Default is it only listens on localhost.

php tort.phar qs

To accept network connections you can bind it to an IP or 0.0.0.0 for to listen to any interface. You are responsible for any firewalling or port forwarding needed.

php tort.phar qs --bind=0.0.0.0

All of the following commands accept --host and --port options to talk to a queue server running on another machine on the other side of the network.

Asking the queue server to perform a new Text-to-Speech just change the gen command you would normally use into qgen and pass everything else just as you normally would to it.

php tort.phar qgen "this is a test." --seed=69

The queue server can then be asked for its status. By default it will just show a summary and a list of what job IDs are currently running. You can have it show the entire queue with --list and see the full job info including the commands it is going to run with --full.

php tort.phar qstatus

The queue server can be asked to terminate itself. By default it will wait until the currently running jobs finish and then it will quit. It can be asked to terminate itself immediately by adding the --now option, which will then add any jobs being interupted back onto the queue stack to start over the next time the server is run. Interupted jobs can be thrown away instead by adding the --abandon option.

php tort.phar qquit

The queue server can be asked to pause itself after the currently running task finishes, therefore sitting idle instead of running any more jobs. It can be asked to resume itself as well.

php tort.phar qpause

php tort.phar qresume

Thanks to the TorToiSe TTS person for all the work of making it work.

Hope they continue with making more compat models or release more info on how, so we can give the voices better diveristy. Like right now if your goal is to do an old-timey radio transatlantic person, or any city accents, gonna have a hard time.

But it still pretty cool.