Upload NOTICE.md

NOTICE.md

@@ -0,0 +1,26 @@
+# Notices & Attribution
+
+## 1. Bodhan AI / AI4Bharat Indic-Speak
+This repository provides quantized (INT8 ONNX and GGUF) derivatives of **Indic-Speak (preview v2)**, developed by **Bodhan AI** and **AI4Bharat** at the Center of Excellence for Artificial Intelligence in Education (supported by the Ministry of Education, Govt. of India).
+- Upstream Model: [bodhan-ai/indic-speak-preview-v2](https://huggingface.co/bodhan-ai/indic-speak-preview-v2)
+- Upstream Commit: `c36d2c49c18d361860a4f5f540b686e066a362bf`
+- License: Indic Open Model License v1.0 (see [LICENSE](LICENSE) and [LICENSE_DEED.md](LICENSE_DEED.md))
+- Scope: Zero-shot, highly expressive multilingual and cross-lingual text-to-speech synthesis supporting 23 Indian languages + English, across 46+ curated speaker voices.
+- Mandatory Attribution Notice: **"Built with Indic-Speak from Bodhan AI / AI4Bharat."**
+
+## 2. Base Architecture Credits
+1. **Autoregressive Acoustic Transformer**: Based on **Meta Llama 3.2 3B** (~3.21B parameters), fine-tuned for generating discrete speech tokens from phonetic text input.
+   - Base Model: [meta-llama/Llama-3.2-3B](https://huggingface.co/meta-llama/Llama-3.2-3B)
+   - Base Model License: [Llama 3.2 Community License Agreement](https://llama.meta.com/llama3/license/)
+2. **Neural Vocoder**: Based on **Vocos** architecture, designed by Charactr Inc.
+   - Upstream Vocos License: Apache 2.0
+   - Function: Reconstructs 24 kHz high-fidelity continuous waveform audio from acoustic token embeddings via inverse Short-Time Fourier Transform (iSTFT) with 257 frequency bins (n_fft=512, hop_length=128).
+
+## 3. Derivative Modifications
+- **GGUF Quantization of Llama 3.2 Backbone**:
+  - `indic-speak-q8_0.gguf` (3.27 GB) and `indic-speak-q5_k_m.gguf` (2.23 GB), reducing memory footprint by up to 66.2% from the 6.6 GB FP32 checkpoint while retaining near-lossless acoustic token probabilities.
+- **ONNX INT8 Quantization of Vocos Neural Vocoder**:
+  - `vocos_backbone_int8.onnx`: Quantized from 222.7 MB FP32 down to **79.8 MB INT8**, achieving ultra-fast **13.3 ms CPU latency** per chunk on standard edge processors.
+  - Native export: Clean graph without operator warnings, compatible with ONNX Runtime CPU and DirectML.
+- **Target Distribution**: Client-side local inference for private voice assistants, reading tools, on-device screen readers, and low-latency interactive audio systems. No multi-tenant hosted API services are provided under this distribution.
+- **Derivative Release**: Packaged and published under the Indic Open Model License v1.0 by Community Contributor (2026).