AI
New multilingual speech synthesis system delivers fast, efficient text-to-speech across 12 scripts including Indian languages
Ninada achieves low-latency speech generation for mobile and web by splitting VITS models into streaming sub-components, overcoming browser and device constraints.
Researchers have unveiled Ninada, an open-source text-to-speech system designed to overcome the technical barriers that prevent real-time speech synthesis on mobile devices and in web browsers. The system synthesizes natural-sounding audio at 24 kHz across twelve languages—English, Sanskrit, Hindi, Kannada, Marathi, Bengali, Tamil, Telugu, Malayalam, Gujarati, Punjabi, and Odia—while maintaining practical speed and memory footprints.
The core problem#
Existing end-to-end VITS (Variational Inference Text-to-Speech) models generate speech in a single pass but struggle with client-side deployment. Standard ONNX exports run entirely within memory before producing audio, causing delays of several seconds before the first sound reaches the listener. Additionally, supporting multiple languages creates compounding challenges: language-specific models can exceed 138 MB each in standard format, overwhelming browser caches on mobile devices. When WebGPU compilation or WebAssembly execution begins, users face stalls of 6-14 seconds just for the system to load, while the actual synthesis often runs slower than real-time playback speed. Audio pipelines on mobile operating systems further complicate matters, with frame drops and muting occurring under resource contention.
How Ninada splits the workload#
Rather than treating speech synthesis as a monolithic operation, Ninada divides a single model into streaming sub-components that process audio incrementally. The system partitions the model into three sections—prior, flow, and vocoder—that operate in a windowed fashion, overlapping audio segments with raised-cosine crossfades to create seamless output. This split-graph approach reduces Time-To-First-Audio (TTFA) significantly and lowers peak memory consumption.
The system deploys two vocoder specialists depending on the device. Desktop and WebGPU systems use HiFi-GAN v1 models (27.9 MB per language in compressed ONNX format), while mobile platforms run distilled iSTFTNet vocoders (21.7 MB) that require 2.64× fewer processing operations through an exact inverse STFT implementation avoiding complex arithmetic. A lightweight Klatt-based fallback adds zero additional file size overhead for minimal audio quality.
Performance metrics and real-world results#
Testing on 480 held-out 24 kHz utterances, Ninada achieved 3.22 dB Mel-Cepstral Distortion (MCD) in the 0-8 kHz band on desktop, with the mobile variant reaching 3.09 dB. By comparison, Meta's much larger MMS system (1,526.8 MB for 16 kHz synthesis) produced 5.17 dB error, demonstrating superior efficiency despite Ninada's leaner design. The system produced audible output within on WebGPU-enabled desktops and maintained zero-dropout streaming at 0.75-0.76× real-time speed on a Google Pixel 10 Pro XL handset, latency measured at 538-649 milliseconds.

