IndicConformer · Sanskrit ASR
100% Client-Side

सुश्रोता Su-śrotā

॥ भद्रं कर्णेभिः शृणुयाम ॥ “May we hear what is auspicious”

Scholar-grade Sanskrit speech recognition powered by IndicConformer-CTC and ONNX Runtime WebAssembly. Audio never leaves your device.

WebML-Kit 115M Conformer-CTC WASM SIMD Multi-threaded INT8 Quantized Local Cache Storage

Initialize Sanskrit ASR Model

Loads the INT8 quantized Conformer-CTC acoustic network (~178 MB) and mel-filterbank preprocessor via WebML-Kit. Persistent browser caching ensures instant startup on return visits.

00:00
Or Try Authentic Sanskrit Recitations
Drop an audio file here or click to browse
Supports .wav, .mp3, .ogg, .m4a (automatically resampled to 16 kHz)
Recent Transcriptions
Transcriptions from this session will appear here.

About Su-śrotā (सुश्रोता)

Su-śrotā is a Conformer-CTC speech recognition model originally trained by Prathosh AP's lab at the Indian Institute of Science (IISc). This web edition runs 100% client-side via WebML-Kit and ONNX Runtime WebAssembly with SIMD acceleration.

The 115M-parameter acoustic network is dynamically quantized to INT8 (~178 MB), and mel-filterbank feature extraction is executed in a dedicated ONNX preprocessor graph. Model weights are cached locally in your browser's Cache Storage for instant offline use.

Model weights are available at huggingface.co/gnumanth/sushrota-sanskrit-asr-onnx. Source code and export scripts are open source on GitHub.