OpenPhonemizer

Audio Samples / Models / Live Demo / Dataset

A permissively licensed, open sourced, local IPA Phonemizer (G2P) powered by deep learning. This Phonemizer attempts to replicate the espeak Phonemizer while remaining permissively-licensed.

OpenPhonemizer is heavily based on the amazing DeepPhonemizer. The main changes are the model checkpoints, which more closely resembles espeak's phonemizer.

Optional GPL-licensed portions are available here.

Features

• Permissively licensed & open source
• Fast & efficient
• Works well with TTS models that depend on phonemizer or espeak
• Automatic GPU acceleration (CUDA/MPS) if available

Project

• Project status: Alpha
• Supported languages: English (more coming soon! What languages do you want? Let me know!)

Installation

Easily install OpenPhonemizer:

pip install -U openphonemizer

Or, install the latest version from Git:

pip install -U "openphonemizer @ git+https://github.com/NeuralVox/OpenPhonemizer"

Usage

OpenPhonemizer

from openphonemizer import OpenPhonemizer
phonemizer = OpenPhonemizer()
# Or specify a custom checkpoint path: OpenPhonemizer('model.pt')
phonemizer('test')
phonemizer('hello this is a test')

Please note that by default, OpenPhonemizer loads a built-in dictionary of words/phonemes. Because storage is quite inefficient, the model is ~100MB larger and uses more memory, however it is much faster. If you're low on VRAM, you can either run the model exclusively on CPU (disable_gpu=True) or load a model without a dictionary.

Load without dictionary:

from cached_path import cached_path
from openphonemizer import OpenPhonemizer
phonemizer = OpenPhonemizer(str(cached_path('hf://openphonemizer/ckpt/best_model_no_optim.pt'))) # add disable_gpu=True to run on CPU only
phonemizer('test')
phonemizer('hello this is a test')

[NEW] Use autoregressive model:

NEW: An autoregressive model is now available. The autoregressive model is more accurate but slightly slower. To use the autoregressive model:

OpenPhonemizer(str(cached_path('hf://openphonemizer/autoreg-ckpt/best_model.pt')))

Evaluation

We introduce PhonemizerBench, a benchmark to evaluate the similarity of alternate Phonemizers to espeak (this benchmark measures against espeak, assuming it's score is 100).

Run 1

Run 2

Run 3

Todo

• Train autoregressive model
• Allow disabling GPU usage
• Multilingual support (any requests?)

License

OpenPhonemizer is open source software. You may use it under the BSD-3-Clause Clear license found in the LICENSE file.

Please note that OpenPhonemizer depends on software under different licenses, it is your responsibility when redistributing or modifying OpenPhonemizer to comply with these licenses (notably LGPL).

By contributing to this repository, you grant the author the permission to change the license in the future at their sole discretion or offer different licenses to other individuals.

NOTE: Model weights may be licensed under different licenses. Please make sure to check all model weights for licenses.

Credits

Special thanks to Christian Schäfer, who created Deep Phonemizer, on which OpenPhonemizer relies. OpenPhonemizer uses num2words to read out large numbers and cached_path from Allen AI for caching models.

OpenPhonemizer was created by mrfakename.