Twitter Thread: link

Deployed Website: tts.themetavoice.xyz

A "pay-as-you-go" API for Tortoise TTS. It uses Modal underneath.

Tortoise is one of the best text-to-speech systems ever built, but it currently requires the user to deploy their own service on a GPU which can be time-consuming, difficult & expensive. The alternative is to use paid services which offer a monthly pricing tier and are closed-off. This repository aims to provide a usage-based, pay-as-you-go API based on open-source code instead.

We have made some improvements to Tortoise to make the inference ~30% faster, and welcome contributions on our repo to improve it further!

We also provide a Python API wrapper that can used for easy integration into your Python code.

You can use the HTTP end-point directly, or the provided Python wrapper.

The synthesis can be done by send a POST request to "https://vatsalaggarwal--tts-app.modal.run".

Make sure to include headers that keep the connection alive sufficiently long as the inference can take 30s-150s.

"Connection: keep-alive",
"Keep-Alive: timeout=600, max=100",

• api_key: Key that manages payment (note the charges are "at-cost"). It can be generated by visiting https://tts.themetavoice.xyz
• text: Text you want to synthesize. Normalized text (1 -> "one") with proper punctuation works best. 225 characters is probably the maximum you should try given how the model was trained, and the constraints on the Modal backend.
• voices: String specifying which voice should be used to synthesize the text. There are four ways to get different voices out of this model:
  • "random": The model randomly picks a voice in its embedding space
  • "<name>": Use one of the voices the model was trained on (e.g. train_grace)
    • Choices are: angie, applejack, cond_latent_example, daniel, deniro, emma, freeman, geralt, halle, jlaw, lj, mol, myself, pat, pat2, rainbow, snakes, tim_reynolds, tom, train_atkins, train_daws, train_dotrice, train_dreams, train_empire, train_grace, train_kennard, train_lescault, train_mouse, weaver, william
  • "<name>&<name>": Combine two voices (e.g. train_grace&emma)
  • "": (Zero-shot) Used with target_file parameter described next to synthesize text in the voice of an utterance given by the user.
• [optional] target_file: A pointer to a mp3 or wav file from a speaker whose voice you're trying to synthesize text in. This should be uploaded to a static file store (like a public s3 bucket).
  • This parameter is unused if voices is specified, and shouldn't be specified as part of the body.
  • If this parameter is used, make sure that voices is set to "".

run_api.py provides an example.

• Use GPT-3 to perform text normalisation
• Inference optimisations
  • TODO: add ideas
• Finetuning code

• python=3.10.8 (important for Modal)