Acknowledgements

We're grateful to:

• Lambda and Stephen Balaban for supporting our model training through their foundational model grants.
• The ONNX Runtime community for building a fast, cross-platform inference engine.
• Alexander Veysov for the great Silero Voice Activity Detector.
• Viktor Kirilov for his fantastic DocTest C++ testing framework.
• Nemanja Trifunovic for his very helpful UTF8 CPP library.
• The Pyannote team for making available their speaker embedding model.
• The espeak-ng community, for all of their inspiring work tackling the endless complexities of translating the written word into speech.
• The CMU Pronouncing Dictionary and eSpeak NG for English G2P lexicon and pronunciation filtering (core/moonshine-tts/data/en_us).
• open-dict-data/ipa-dict for multilingual IPA lexicon data used across many locales (core/moonshine-tts/data).
• WikiPron (CUNY-CL) for Italian, Russian, and European Portuguese pronunciations.
• Koichi Yasuoka for the Hugging Face models chinese-roberta-base-upos, roberta-small-japanese-char-luw-upos, and roberta-base-korean-morph-upos.
• hexgrad/Kokoro-82M and onnx-community/Kokoro-82M-ONNX for Kokoro TTS weights and ONNX (core/moonshine-tts/data/kokoro).
• PiperTTS for their excellent lightweight TTS models.
• MeloTTS from MyShell as reference for Korean Piper voice training (core/moonshine-tts/data/ko).
• English Wiktionary and hermitdave/FrequencyWords for Hindi lexicon material (core/moonshine-tts/data/hi).
• hbenbel/French-Dictionary for related French liaison lexicon work (core/moonshine-tts/data/fr).
• AbderrahmanSkiredj1/arabertv02_tashkeel_fadel for Arabic diacritization and CAMeL Tools for optional Arabic MSA lexicon builds (core/moonshine-tts/data/ar_msa).
• ZipVoice for their high-quality text to speech and voice cloning.
• The team behind the VCTK dataset at the University of Edinburgh for generously providing a rich source of voice styles.