| --- |
| language: |
| - sa |
| - hi |
| - kn |
| - ta |
| - te |
| - ml |
| license: mit |
| library_name: onnx |
| pipeline_tag: text-to-speech |
| tags: |
| - text-to-speech |
| - sanskrit |
| - indic |
| - vits |
| - hifi-gan |
| - onnxruntime-web |
| - wasm |
| --- |
| |
| # `sntts` — Pan-Indic & Classical Sanskrit Neural TTS (VITS + 16 kHz HiFi-GAN) |
|
|
| `sntts` is a compact, single-checkpoint **Pan-Indic & Classical Sanskrit** neural text-to-speech model (`35.4M` parameters; **`54.3 MB` single-file browser ONNX**) supporting **6 Indic languages** (`Sanskrit [sa]`, `Hindi [hi]`, `Kannada [kn]`, `Tamil [ta]`, `Telugu [te]`, `Malayalam [ml]`) across **9 Brahmic scripts + ISO-15919/IAST** in Python, Node.js, and 100% client-side in the browser via `onnxruntime-web` WASM SIMD. |
|
|
| ## Training & Evaluation Highlights |
| - **GCP GPU Training**: Trained on `NVIDIA L4` (`24` epochs, `3024` steps) across `1,008` studio 16 kHz utterances curated from 7 Hugging Face speech datasets (`prathoshap/sushrota-sanskrit-asr-data`, `yashnbx/gita_supersite_sanskrit_tts`, SYSPIN Hindi/Kannada/Tamil/Telugu, and IITM Malayalam). |
| - **Held-Out Evaluation (`N = 126` utterances, `B = 1,000` bootstrap)**: |
| - Mean Mel-Cepstral Distortion (`MCD`): **5.143 dB** |
| - Independent `Su-śrotā` Sanskrit ASR Round-Trip `CER`: **15.006%** |
| - **Files**: |
| - `model.safetensors`, `config.json`, `sntts_config.json`, `vocab.json` — PyTorch FP32 checkpoint |
| - `onnx/model_quantized.onnx` — Single-file FP16 browser ONNX (`54.3 MB`, zero `.data` sidecars, `r = 0.9998` parity with FP32) |
| - `training_history.json`, `gcp_job_manifest.json` — Full training telemetry |
|
|
