New research allows deployment of speech AI models directly onto client devices for Indic languages. This eliminates cloud transcription costs, network latency, and server-side audio retention. Two new papers, “Swanada” for automatic speech recognition (ASR) and “Ninada” for text-to-speech (TTS), both authored by Hemanth HM, detail open-source systems designed for browsers and mobile runtimes across a dozen scripts.

The core problem for on-device speech AI has been model size and computational demands. Previous monolithic per-language Conformer checkpoints for ASR, for instance, required approximately 177 MiB per script. Across twelve scripts, this totals over 2.1 GiB, far exceeding typical browser CacheStorage quotas. Similarly, VITS text-to-speech models, while powerful, often produce 138.8 MB float32 checkpoints per language, straining browser caches and leading to significant delays. This scale makes client-side deployment impractical for a broad range of Indic languages.

Swanada: Sub-30 MB ASR Across Twelve Scripts

The “Swanada: Split-Encoder Capability-Routed Conformer CTC” paper, submitted on November 10, 2026, presents a 16 kHz streaming ASR system that reduces model sizes to sub-30 MB. This is a significant reduction from the typical 177 MiB per script. Swanada covers Classical Sanskrit, Indian English, and ten Indian languages (Sanskrit, English, Hindi, Kannada, Marathi, Bengali, Tamil, Telugu, Malayalam, Gujarati, Punjabi, Odia).

Hemanth HM identified four system bottlenecks for neural exports in browsers and mobile runtimes. Swanada addresses them:

  • Monolithic per-language Conformer checkpoints: Swanada uses a split-encoder and capability-routed architecture to avoid the large per-script model sizes.
  • Static relative positional encoding tables: These tables previously embedded 10.2-20.5 MB of FP32 constants. Swanada optimizes this.
  • Slow or unsupported ConvInteger kernels: 1×1 Conv1d layers frequently triggered these issues in WebAssembly SIMD.
  • Sliding-window CTC flickering and mobile WebAudio pipeline issues: These caused problems with boundary tokens and silent stalls under iOS WKWebView.

By tackling these issues, Swanada makes it feasible to run high-quality ASR for multiple Indic languages directly on a user’s device. This bypasses the need to send audio to a cloud server, which can reduce latency and improve data privacy.

Ninada: Streaming TTS with Reduced Latency

Complementing the ASR work, the “Ninada: Sub-Clause Split-Graph VITS Streaming” paper, submitted on November 5, 2026, introduces a 24 kHz TTS system. This system overcomes the latency and memory spikes associated with end-to-end VITS models, which typically execute monolithically over an entire clause, delaying the Time-To-First-Audio (TTFA) by several seconds.

Ninada also addresses four client-side deployment bottlenecks for Indic scripts:

  • Large checkpoints: The system moves away from 138.8 MB float32 checkpoints per language, instead using a capability-routed 48-checkpoint fleet that includes 24 studio HiFi-GAN v1 specialists, each weighing 27.9 MB.
  • Mobile WebGPU stalls: Previous models faced 6-14 second stalls during convolution shader compilation, often running slower than real time (1.44 times RTF).
  • WebAssembly pthread pool stalls: Small pthread pools on big-LITTLE mobile CPUs caused stalls at barriers.
  • Web Audio pipeline issues: Mobile pipelines dropped frames or muted output under AudioWorklet contention and iOS WKWebView rules.

The Ninada system covers English, Classical Sanskrit, and ten Indian languages (Sanskrit, English, Hindi, Kannada, Marathi, Bengali, Tamil, Telugu, Malayalam, Gujarati, Punjabi, Odia), matching the language coverage of Swanada. Sub-clause split-graph VITS streaming achieves faster TTFA and reduces activation memory. This makes speech synthesis feel more responsive to the user.

Implications for Indic Language AI Development

These two papers from Hemanth HM build a practical, full-stack speech AI solution for Indic languages that operates entirely on the client side. While previous research, such as the SraVaani-1.0 ASR model, has focused on broader language coverage (65 Indian languages and dialects), Swanada and Ninada prioritize the system-level optimizations needed to make a core set of 12 Indic scripts viable for on-device use. This shift in focus means that developers can now consider building applications that offer immediate speech recognition and synthesis without the overhead of cloud infrastructure.

Addressing specific WebAssembly, WebGPU, and Web Audio pipeline limitations directly impacts developers working on browser-based or mobile applications for Indian users. For instance, our coverage on “Indic Language AI: Tokenizer Efficiency and Tamil-First Models Advance Research” showed the need to optimize individual components. Swanada and Ninada extend this by optimizing the entire model architecture and runtime for specific client environments. The practical consequence is a reduction in operational costs for companies, improved user experience due to lower latency, and improved privacy, as audio data never leaves the user’s device for processing. This move away from cloud reliance could open up new application categories for Indic language users, especially in areas with intermittent internet connectivity or strict data privacy requirements.

Sources

Compiled by Launch91 Desk from the sources linked above. More about Launch91.