创建

登录 ReadmeX

或

了解我们如何处理你的信息: 隐私政策

ReadmeX

陪你读懂 AI 动态

今天,想先了解什么?

选一份动态总结,或直接聊聊你的问题。

社区和关注动态总结只属于你自己。

资讯

Ninada v5.1.0开源多语种浏览器TTS

AI 总结

开发者 Hemanth 宣布开源 Ninada v5.1.0,其中代码采用 MIT 许可证,48 个 24 kHz ONNX 文本转语音检查点采用 CC-BY-4.0 许可证。该版本覆盖古典梵语、英语和 10 种印度语言,并提供无需注册、完全在本地运行的浏览器工作室。开发者还报告称,手机 WASM 合成延迟为 538 至 649 毫秒,而 Mobile WebGPU 在 Pixel 10 Pro XL 测试中表现更差。

深度见解

Ninada 以开放代码、检查点和本地浏览器推理提供多语种文本转语音支持,其中包含古典梵语和多种印度语言。

Building an in-browser TTS model for Indic languages, including Classical Sanskrit, has been a longtime dream of mine. Today I'm open-sourcing Ninada (code MIT, 48 ONNX checkpoints CC-BY-4.0).

Live in-browser studio (zero sign-up, runs 100% locally): https://h3manth.com/ai/ninada/

Why I built it (and what broke on real phones)

Most open Indic TTS setups either rely on cloud APIs, ship multi-gigabyte unified checkpoints that bleed accents across scripts, or use 16 kHz Meta MMS VITS (1.5 GB across 12 languages, 5.17 dB MCD) exported as a monolithic ONNX graph that blocks until the entire sentence finishes.

When I profiled streaming VITS on a physical Pixel 10 Pro XL over wireless ADB, two common browser ML assumptions failed immediately: 1. Mobile WebGPU lost to WASM: Dawn shader compilation stalled cold start for 6 to 14 seconds, and GPU buffer readback pushed synthesis slower than real time (1.48x RTF). 2. 2-3 WASM threads ran slower than 1 thread: Heterogeneous big.LITTLE core scheduling stalled ONNX Runtime spin barriers (1.76x RTF at 2 threads vs 1.19x at 1 thread), only recovering at 4 threads.

#Ninada v5.1.0#ONNX Runtime#VITS#iSTFTNet#HiFi-GAN#WebGPU#WASM#Cascade Klatt-Akshara

原文约 2 分钟读完阅读原文
群聊

还没有评论,来说说你的看法。