Misaki is a versatile grapheme-to-phoneme (G2P) engine designed for speech synthesis models, supporting English, Japanese, Korean, Chinese, and Vietnamese. It offers multiple tokenization strategies and fallback mechanisms for out-of-vocabulary words, catering to researchers and developers in the speech technology domain.

How It Works

Misaki employs a modular design, with distinct tokenizers and phonemizers for each supported language. For English, it offers a transformer-based approach (when trf=True) leveraging contextual word embeddings for homograph disambiguation, and a non-transformer fallback to espeak-ng for robustness. Japanese tokenization utilizes pyopenjtalk with unidic for pitch accent, while older versions relied on cutlet, fugashi, and mecab. Korean and Chinese tokenization are adapted from established libraries like g2pK and paddlespeech, respectively.

Quick Start & Requirements

• Install English support: !pip install -q "misaki[en]"
• For espeak-ng fallback: !apt-get -qq -y install espeak-ng and !pip install -q "misaki[en]" phonemizer-fork
• Official demo: https://hf.co/spaces/hexgrad/Misaki-G2P

Highlighted Details

• Supports multiple languages: English, Japanese, Korean, Chinese, Vietnamese.
• Offers transformer-based English G2P with contextual word embeddings for homographs.
• Integrates pyopenjtalk and unidic for Japanese pitch accent.
• Leverages g2pK and paddlespeech for Korean and Chinese tokenization.

Maintenance & Community

• Active development with contributions from various individuals and projects.
• Discord server available: https://discord.gg/QuGxSWBfQy

Licensing & Compatibility

• The README does not explicitly state a license. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The project is still under active development with several items on the TODO list, including data compression and training seq2seq fallback models. The handling of homographs using BERT and logistic regression is noted as an area for escalation.