presenting swar स्वर 🇳🇵
nepali has about 32 million speakers, yet almost no open text-to-speech sounds like someone from nepal.
we built swar, an open nepali voice model. give it 20 to 30 seconds of anyone's voice & it speaks for them: nepali, english, and everything in between.
on NepTTS-Bench, swar has the lowest pronunciation error: 2.92% beating omnivoice at 4.32%, elevenlabs v3 is at 12.91%.
it's #1 on the nepali voice cloning benchmark (fleurs, clean track), and makes half the errors on nepali mixed with english: 5.6 vs 11.4.
the code, weights, benchmark and a free colab notebook are all open. runs on your own gpu.
only clone a voice with that person's permission.
github: t.co/DpLOXFBa0o
hugging face: t.co/QLrfPT7r9q
website:
Art, video & music
Open Nepali voice model beats ElevenLabs on pronunciation
Arjan, a 16-year-old builder, and their team created swar, an open-source Nepali voice cloning model. It achieves 2.92% pronunciation error on NepTTS-Bench, surpassing OmniVoice and ElevenLabs v3, with code, weights and a Colab notebook.
506likes
61reposts
262saves
27Kviews
- Role
- 16-year-old builder
- Posted
