PaperVoice Models (v2.0.0)
Official model repository for PaperVoice — the open-source, 100% offline CPU research PDF reader with read-along word highlighting and mathematical formula narration.
PaperVoice automatically downloads these model weights on first start and verifies every file against the cryptographic SHA-256 checksums below.
Model Inventory & Components
1. Document Layout & OCR Engine
layout_heron/(Docling Heron RT-DETR): Page object detection (paragraphs, section headers, tables, display math). Converted to ONNX with pure NumPy preprocessing.ocr/(PP-OCRv6 small + direction classifier via RapidOCR): High-efficiency text detection and recognition for scanned research papers.
2. Mathematics & Table Structure Recognition
ppfn_plus_s.onnx(PP-FormulaNet+ S): Recognizes mathematical formulas directly from page crops into LaTeX (~58M parameters, ~50× faster on CPU than CodeFormulaV2).tableformer_fast/(TableFormer Fast): Recovers row-and-column grid structure and cell coordinates from complex academic tables.
3. Neural Voice Engine (v2.0 Kokoro-82M)
kokoro/kokoro-v1.0.onnx(StyleTTS 2 architecture, 82M parameters, 24 kHz): Pristine reference neural voice delivering natural human prosody and phase fidelity ($r = 1.000000$).kokoro/voices-v1.0.bin: Style embedding vectors for voice personalities (af_heart, etc.).kokoro/kokoro-v1.0.fast.onnx: Conv1D STFT-accelerated model (378 ms faster per sentence).kokoro/kokoro-v1.0.fused_snake.onnx: Optimized model with 48 fused Snake periodic activations (1.99 s sentence latency, RTF 0.313).kokoro/snake_op.dll: Zero-dependency statically linked AVX2/FMA C++ custom operator for ONNX Runtime on Windows.
(Note: The v1 classic voice, Piper en_US-lessac-medium, is licensed for research-only and downloaded on demand from Piper's official repository).
Cryptographic Checksums (SHA-256)
Every file downloaded by PaperVoice is validated against these exact checksums:
| File Path | Size | SHA-256 Checksum |
|---|---|---|
| Neural Voice (Kokoro-82M) | ||
kokoro/kokoro-v1.0.onnx |
325.5 MB | 7d5df8ecf7d4b1878015a32686053fd0eebe2bc377234608764cc0ef3636a6c5 |
kokoro/voices-v1.0.bin |
28.2 MB | bca610b8308e8d99f32e6fe4197e7ec01679264efed0cac9140fe9c29f1fbf7d |
kokoro/kokoro-v1.0.fast.onnx |
325.5 MB | bd26413cf0b1505c98b37b392fcb82ad16dbf20d81eddbbaa9f24a5a5c50b573 |
kokoro/kokoro-v1.0.fused_snake.onnx |
325.5 MB | 1abf477b767d56639f154c7a66ff3b55baa32454e043e688137403ad027b94f1 |
kokoro/snake_op.dll |
359 KB | 568e4c9fb3419703a06dc5c71cb4011d6eba4b1d1571c5bb9747ded1d6327660 |
| Formula Recognition | ||
ppfn_plus_s.onnx |
318.4 MB | 8f85ff062a32f137e07dae64d6f955654333b3c0d59e74d2b728ae3907c309a9 |
ppfn_plus_s.tokenizer.json |
1.5 MB | 352f9ea33f61cca2cc5d44bfa11ead41f243c4be96255f39bae9c0246787dc9d |
| Page Layout (Heron) | ||
layout_heron/model.onnx |
114.6 MB | 59c81a3a2923042d85034ffc487f8f47e4854117e879aef89b2b9f728fb4922a |
layout_heron/config.json |
3 KB | c6e67b6cdf64fa245779d0409bb994aa96e8c1790e15ec07a65843628e232180 |
layout_heron/preprocessor_config.json |
422 B | 54d086cf0d7d371f7fac36e5d3a0dae31e211298affc816106cc376506799d58 |
| Table Structure (TableFormer Fast) | ||
tableformer_fast/encoder.onnx |
89.2 MB | f8f871b0294f456c43761d503af85a0a5abf4132c705ef6f5a7ba32dd505bb26 |
tableformer_fast/tag_step.onnx |
42.6 MB | 8e1856360d6ea2156cd756969915046a5d1e4744b92fcc507ffe921efe48fc37 |
tableformer_fast/bbox.onnx |
42.7 MB | 0315f1cffd7acbc3a831b2de5bd44cfc1319265aafcb5d0d08f50b790d02a89a |
tableformer_fast/tm_config.json |
3 KB | dca6762508dddfae6d57d6cb4ef822c6000119dff0f3b6489db7413118c2622a |
| Scanned Document OCR (RapidOCR) | ||
ocr/PP-OCRv6_det_small.onnx |
4.6 MB | 090f04abcd9d9a7498bc4ebf677e4cb9bdce1fe4197ddb7e529f1ef44e1ff94f |
ocr/PP-OCRv6_rec_small.onnx |
16.1 MB | 6f327246b50388f3c176ae304bd95767ea6dc0c9ae92153ef8cbe210b3c14884 |
ocr/ch_ppocr_mobile_v2.0_cls_mobile.onnx |
1.4 MB | e47acedf663230f8863ff1ab0e64dd2d82b838fceb5957146dab185a89d6215c |
ocr/ppocrv6_dict.txt |
114 KB | b5f2bfe2bdd9448429e3e82b51c789775d9b42f2403d082b00662eb77e401c5d |
Performance on Laptop CPU (Intel Core i5, 4 Cores)
- Time-to-First-Sound (TTFS): 3.28 s – 4.79 s (streaming 130-char clause chunks).
- Real-Time Factor (RTF): 0.313 – 0.447 ($>2.2 imes$ faster than real-time speech).
- CPU Utilization: 25% (2 dedicated intra-op threads, eliminating cache false sharing).
- Acoustic Fidelity: $r = 1.000000$ waveform correlation, SNR $> 95 ext{ dB}$.
Upstream Licenses & Attribution
- Kokoro-82M: hexgrad/Kokoro-82M (Apache-2.0)
- Docling Heron: docling-project/docling-layout-heron-onnx (Apache-2.0)
- PP-FormulaNet+ S & PP-OCR: PaddlePaddle/PaddleOCR (Apache-2.0)
- TableFormer: docling-project/docling-models (CDLA-Permissive-2.0)