॥ वाचं धेनुमुपासीत ॥
A 100% client-side, vṛtta-aware Sanskrit śloka-to-chant neural synthesis engine. Runs a 22-block Flow-Matching DiT, Fourier-Hann Vocos vocoder, and Pāṇinian prosody analyzer directly on your device via WebGPU and WebAssembly SIMD—with zero server calls.
Synthesized 100% locally in your browser tab from Devanagari, Kannada, Telugu, Malayalam, Tamil, or Bengali.
Zero-latency Guru (ˉ) and Laghu (˘) syllable weight analysis computed in JavaScript.
Prevents schwa-deletion while preserving jihvāmūlīya, upadhmānīya, and retroflex clusters before ONNX tokenization.
How a 1.8 GB Sanskrit Flow-Matching TTS pipeline was baked into a 237 MB WebGPU and WebAssembly runtime with 0.9999 backbone cosine fidelity.
Instead of running server-side audio slicing and STFT extraction at runtime, all 18 classical meter templates and 3 alliteration repeat-primes are pre-baked into a 2.45 MB IEEE-754 Float16 binary (baked_bank.bin) decoded directly into Float32Array tensors in JavaScript.
Factoring all 23 time-conditioning projections onto the 32-D trajectory of silu(time_embed(t)) eliminates 137.5M parameters. Combined with per-column symmetric INT8 storage, fused QKV projections, and exact Head-0 split-half RoPE, the 22-block DiT step achieves 0.99992 cosine similarity inside Chrome WebGPU.
vagdhenu-text.js implements zero-dependency multi-script detection, SLP1 normalization, Guru/Laghu syllable weight tagging, 18-vṛtta pattern matching, visarga sandhi, satva assimilation, and Kannada orthographic routing entirely on the client thread.
WebGPU and ONNX lack native Complex64 inverse STFT ops. We reformulate torch.istft as an exact 1026-channel Fourier-Hann ConvTranspose1d basis paired with an unquantized FP32 instantaneous phase head, followed by an in-browser fricative/halant RMS gate and 24 kHz PCM WAV encoder.
Hosted at huggingface.co/gnumanth/sanskrit-tts-onnx and cached locally in your browser's Cache API for offline-capable, zero-server Sanskrit recitation.
vagdhenu_cond_q8.onnx · 34 MB
Hoists the 4-block ConvNeXtV2 text encoder, reference mel projection (W_c @ cond_mel + b), text projection (W_t @ emb), and split-half rotary position embeddings outside the ODE loop so they execute only once per hemistich instead of at every flow step.
vagdhenu_dit_step_q8.onnx · 192 MB
Executes sway-sampled Rectified Flow ODE integration with Adaptive Classifier-Free Guidance directly on the GPU via WebGPU compute shaders (with automatic fallback to WebAssembly SIMD). Uses single-node FastGelu, static head reshapes, and Rank-32 SVD AdaLN.
vagdhenu_vocos_q8.onnx · 56 MB)
Download from Hugging Face →
baked_bank.bin · 2.84 MB)
Browse Model Repository →