Dev open-sources Ninada: 48 on-device 24kHz TTS voices across Sanskrit and 12 languages
init0 · reddit · 2026-11-06
The developer open-sources Ninada (MIT code, 48 ONNX checkpoints CC-BY-4.0): a sign-up-free, fully in-browser 24 kHz TTS covering Classical Sanskrit, English and 10 Indic languages with 48 male/female VITS specialist voices.
Engineering highlights:
- Device-aware: desktops run a 27.9 MB HiFi-GAN (89 ms WebGPU); phones auto-switch to a 21.7 MB distilled iSTFTNet (2.64x lighter)
- Sub-second streaming: VITS ONNX split in-browser in <15 ms, cutting browser RAM 72% to 162 MB; speech in 538 ms on phones with zero dropouts
- Context: Meta's 16 kHz MMS models take 1.5 GB, and mobile WebGPU froze 6-14 s compiling shaders
- Quality: 3.09–3.22 dB MCD across 480 clips, offline caching, 0 MB Klatt fallback
Available via npm/pip with a live in-browser studio.
Related event: Ninada: Open-Source Browser-Based TTS for Sanskrit and 11 More Languages(2 posts)→
More from Multimodal
- SphereEncoder2 hits SOTA pixel-space image quality with single forward pass — tomgoldsteincs · 2026-11-06
- Suno teases that v6 comes in three versions — suno · 2026-11-06
- Community-built Krea-2-Turbo v2 Gradio Space trends on Hugging Face — Intruzo · 2026-11-06
- Alibaba's TaoMate-HD streams high-res audio-video via progressive-resolution autoregression — TaoLiveAIGC · 2026-11-06
- 7-person team Krea ranks #3 in Multi-Image Edit Arena, CEO discusses training and open-source plans — arena · 2026-11-06
- DRAMA expression-editing model launches free on Hugging Face for video and image editing — SimplyAnnisa · 2026-11-06