Tortoise is primarily an autoregressive decoder model combined with a diffusion model. Both of these have a lot of knobs that can be turned that I've abstracted away for the sake of ease of use. I did this by generating thousands of clips using various permutations of the settings and using a metric for voice realism and intelligibility to measure their effects. I've set the defaults to the best overall settings I was able to find. For specific use-cases, it might be effective to play with these settings (and it's very likely that I missed something!)
These settings are not available in the normal scripts packaged with Tortoise. They are available, however, in the API. See
api.tts for a full list.
Some people have discovered that it is possible to do prompt engineering with Tortoise! For example, you can evoke emotion by including things like "I am really sad," before your text. I've built an automated redaction system that you can use to take advantage of this. It works by attempting to redact any text in the prompt surrounded by brackets. For example, the prompt "[I am really sad,] Please feed me." will only speak the words "Please feed me" (with a sad tonality).
Tortoise ingests reference clips by feeding them through individually through a small submodel that produces a point latent, then taking the mean of all of the produced latents. The experimentation I have done has indicated that these point latents are quite expressive, affecting everything from tone to speaking rate to speech abnormalities.
This lends itself to some neat tricks. For example, you can combine feed two different voices to tortoise and it will output what it thinks the "average" of those two voices sounds like.
Use the script get_conditioning_latents.py to extract conditioning latents for a voice you have installed. This script
will dump the latents to a .pth pickle file. The file will contain a single tuple, (autoregressive_latent, diffusion_latent).
Alternatively, use the api.TextToSpeech.get_conditioning_latents() to fetch the latents.
After you've played with them, you can use them to generate speech by creating a subdirectory in voices/ with a single ".pth" file containing the pickled conditioning latents as a tuple (autoregressive_latent, diffusion_latent).
Out of concerns that this model might be misused, I've built a classifier that tells the likelihood that an audio clip came from Tortoise.
This classifier can be run on any computer, usage is as follows:
python tortoise/is_this_from_tortoise.py --clip=<path_to_suspicious_audio_file>
This model has 100% accuracy on the contents of the results/ and voices/ folders in this repo. Still, treat this classifier as a "strong signal". Classifiers can be fooled and it is likewise not impossible for this classifier to exhibit false positives.
Tortoise TTS is inspired by OpenAI's DALLE, applied to speech data and using a better decoder. It is made up of 5 separate models that work together. I've assembled a write-up of the system architecture here: https://nonint.com/2022/04/25/tortoise-architectural-design-doc/
These models were trained on my "homelab" server with 8 RTX 3090s over the course of several months. They were trained on a dataset consisting of ~50k hours of speech data, most of which was transcribed by ocotillo. Training was done on my own DLAS trainer.
I currently do not have plans to release the training configurations or methodology. See the next section..
Tortoise v2 works considerably better than I had planned. When I began hearing some of the outputs of the last few versions, I began wondering whether or not I had an ethically unsound project on my hands. The ways in which a voice-cloning text-to-speech system could be misused are many. It doesn't take much creativity to think up how.
After some thought, I have decided to go forward with releasing this. Following are the reasons for this choice:
- It is primarily good at reading books and speaking poetry. Other forms of speech do not work well.
- It was trained on a dataset which does not have the voices of public figures. While it will attempt to mimic these voices if they are provided as references, it does not do so in such a way that most humans would be fooled.
- The above points could likely be resolved by scaling up the model and the dataset. For this reason, I am currently withholding details on how I trained the model, pending community feedback.
- I am releasing a separate classifier model which will tell you whether a given audio clip was generated by Tortoise or not. See
tortoise-detectabove. - If I, a tinkerer with a BS in computer science with a ~$15k computer can build this, then any motivated corporation or state can as well. I would prefer that it be in the open and everyone know the kinds of things ML can do.
The diversity expressed by ML models is strongly tied to the datasets they were trained on.
Tortoise was trained primarily on a dataset consisting of audiobooks. I made no effort to balance diversity in this dataset. For this reason, Tortoise will be particularly poor at generating the voices of minorities or of people who speak with strong accents.
Tortoise v2 is about as good as I think I can do in the TTS world with the resources I have access to. A phenomenon that happens when training very large models is that as parameter count increases, the communication bandwidth needed to support distributed training of the model increases multiplicatively. On enterprise-grade hardware, this is not an issue: GPUs are attached together with exceptionally wide buses that can accommodate this bandwidth. I cannot afford enterprise hardware, though, so I am stuck.
I want to mention here that I think Tortoise could be a lot better. The three major components of Tortoise are either vanilla Transformer Encoder stacks or Decoder stacks. Both of these types of models have a rich experimental history with scaling in the NLP realm. I see no reason to believe that the same is not true of TTS.