OPEN LINK ↗
// 73d agoOPENSOURCE RELEASE
SILMA AI drops open-source Arabic-English TTS model
SILMA AI has released the weights of its 150M-parameter bilingual text-to-speech model under Apache 2.0, supporting both Modern Standard Arabic and English with voice cloning from under 8 seconds of audio. Built on the F5-TTS diffusion architecture and trained from scratch on tens of thousands of hours of audio, it runs at RTF ~0.12 on an RTX 4090.
// ANALYSIS
Arabic is a vastly underserved language in open-source TTS — SILMA filling that gap with a commercially-permissive, production-ready release is a meaningful contribution to the ecosystem.
- –At 150M parameters, it's roughly half the size of the base F5-TTS config while retaining voice cloning and full Arabic diacritization (Tashkeel) support — a real engineering win for edge/low-resource deployments
- –RTF of ~0.12 makes real-time streaming viable, putting it competitive with ElevenLabs and PlayHT on latency even on a single consumer GPU
- –Apache 2.0 license removes the friction that hampered F5-TTS adoption in commercial products; developers can ship without legal review overhead
- –Arabic NLP has lagged behind English by years; a high-quality open bilingual TTS directly unblocks voice assistants, accessibility tools, and content generation for 400M+ Arabic speakers
- –Available via `pip install silma-tts` with a one-command demo app — unusually smooth DX for an academic-style model drop
// TAGS
silma-ttsspeechaudio-genopen-sourceopen-weights
DISCOVERED
73d ago
2026-03-15
PUBLISHED
73d ago
2026-03-15
RELEVANCE
7/ 10
AUTHOR
oudak2019