cs.SDNov 10, 2026

Swanada: Split-Encoder Capability-Routed Conformer CTC and Sub-30 MB On-Device Speech Recognition Across Twelve Scripts

Authors: Hemanth HM

Abstract

Client-side automatic speech recognition (ASR) across Indic and Classical scripts eliminates cloud transcription costs, network tail latency, and server-side audio retention, yet neural exports face four systems bottlenecks in browsers and mobile runtimes: (1) monolithic per-language Conformer checkpoints require ~177 MiB per script (> 2.1 GiB across 12 scripts), exhausting CacheStorage quotas; (2) static relative positional encoding tables embed 10.2-20.5 MB of FP32 constants while 1x1 Conv1d layers trigger slow or unsupported ConvInteger kernels in WebAssembly SIMD; (3) sliding-window CTC flickers on boundary tokens; and (4) mobile WebAudio pipelines suffer 44.1/48 kHz sample-rate mismatches and silent AudioWorklet stalls under iOS WKWebView. We present Swanada, an open-source 16 kHz streaming ASR system spanning Classical Sanskrit, Indian English, and ten Indian languages (sa, en, hi, kn, mr, bn, ta, te, ml, gu, pa, or) across browser WebAssembly, Node.js, and Python. Swanada pairs capability-routed shared INT8 Conformer encoders (a 17-layer 177.38 MiB Studio tier at d=512d = 512 for desktop/servers and a distilled 14-layer 25.35 MiB Mobile tier at d=256d = 256 for phones) with twelve hot-swappable 515.0 KiB native-script linear CTC heads (257×512257 \times 512), cutting 12-language storage by 11.6×11.6\times (183.4 MiB) and 67.8×67.8\times (31.39 MiB; 25.85 MiB first load). Runtime protobuf tail surgery exposes the shared representation in <2< 2 ms, while dynamic positional synthesis, MatMul pointwise convolutions, and post-quantization QInt8 head calibration preserve sub-30 MB accuracy. Across N=480N = 480 held-out and N=1,280N = 1,280 open public test clips (FLEURS, LibriSpeech, GLOBE, AMI, wild Sanskrit), Swanada achieves 5.43% held-out and 8.87% 12-language open CER on Studio (18.56% and 19.23% on Mobile) at 0.0097×0.0097\times-0.0264×0.0264\times RTF, with 30.6 ms median TTFT and 0.00% committed-prefix churn.

Figures & tables

Explore similar work

CardsList
  1. Typhoon ASR Streaming: Steerable Low-Latency Thai Speech Recognition with Real-Time Shallow Fusion

    Sep 14, 2026Warit Sirichotedumrong, Tanawin Samutsin, Shah Faisal Wani +2Streaming ASRAutomatic Speech Recognition

  2. Design of the IBM Granite 5.0 TurboCTC ASR Model

    Sep 17, 2026Brian Kingsbury, George Saon, Masayuki Suzuki +5TransformerConnectionist Temporal Classification