penra-models / README.md
Likennyy's picture
README: EG2 section heading
69a7981 verified
|
Raw History Blame Contribute Delete
5.94 kB
---
license: other
license_name: mixed-apache-2.0-mit-openmdw-1.1-and-cc-by-sa-4.0
license_link: LICENSE
tags:
- penra
- litert
- on-device
- handwriting-recognition
- speech-recognition
---
# Penra offline models
The model files the [Penra](https://play.google.com/store/apps/details?id=com.penra.app) note app downloads to
read handwriting, write speech and search notes **on the device, with no server**. They are repackaged or converted
from the open models below; the app picks one reading bundle by the phone's memory.
| File | What it is | Made from | License |
|---|---|---|---|
| `penra-ocr-1120.litertlm` | Handwriting reading and speech, phones with 8 GB memory | Gemma 4 E4B, LiteRT-LM bundle with a 1120-token vision signature, trimmed vocabulary | Apache 2.0 (Gemma 4) |
| `penra-ocr-560.litertlm` | The same, faster on phones with a slow GPU (560-token vision signature, MTP drafter for speculative decoding) | Gemma 4 E4B | Apache 2.0 (Gemma 4) |
| `penra-e2b-1120.litertlm` | The same for phones with less memory | Gemma 4 E2B | Apache 2.0 (Gemma 4) |
| `embeddinggemma-2-text-270m.litertlm` | Finding notes by meaning | EmbeddingGemma 2 | Apache 2.0 |
| `uvdoc.onnx` | Flattening a photographed page before reading | UVDoc (PaddlePaddle) | Apache 2.0 |
| `whisper_tr_enc30_dwi4.tflite` | Speech to text, high accuracy (all 15 languages): audio encoder, 30-s window, int4 weights | OpenAI Whisper large-v3-turbo + Penra Turkish LoRA (merged) | MIT (Whisper) |
| `whisper_tr_dec30_kv_mixd.tflite` | The same: text decoder with KV cache, mixed precision | OpenAI Whisper large-v3-turbo + Penra Turkish LoRA (merged) | MIT (Whisper) |
| `whisper_tr_tokenizer.json` | The same: tokenizer | OpenAI Whisper large-v3-turbo | MIT (Whisper) |
| `eg2_tr_fc8.tflite` | Faster Turkish speech to text (Pro): EmbeddingGemma 2 encoder + CTC reader, int8 | [embeddinggemma2-turkish-stt](https://huggingface.co/mstkyvz/embeddinggemma2-turkish-stt) by mstkyvz (on EmbeddingGemma 2) | CC BY-SA 4.0 |
| `eg2_tr_3gram_spoken.bin`, `eg2_tr_unigrams.txt` | The same: spoken-Turkish 3-gram language model and its word list | embeddinggemma2-turkish-stt by mstkyvz (`lm_spoken/`) | CC BY-SA 4.0 |
| `nemotron-asr-step-dwi8.tflite` | Live text while recording: streaming encoder step, int8 | NVIDIA Nemotron 3.5 ASR Streaming 0.6B | OpenMDW 1.1 |
| `nemotron-decoder.tflite` | Live text while recording: RNN-T prediction network and joint | NVIDIA Nemotron 3.5 ASR Streaming 0.6B | OpenMDW 1.1 |
| `nemotron-3-diarization-q8_0.gguf` | Who spoke when (planned) | NVIDIA Nemotron 3 Diarization | OpenMDW 1.1 |
| `wordlists/<lang>/<lang>_50k.txt` | Word lists for the reading check | FrequencyWords (Hermit Dave), commit 525f9b5 | CC BY-SA 4.0 |
`manifest.json` lists every file with its size and SHA-256; the app checks both before using a file.
## Licenses
- Gemma 4 and EmbeddingGemma 2: Apache License 2.0 ([Gemma 4 license](https://ai.google.dev/gemma/docs/gemma_4_license)).
- NVIDIA Nemotron 3.5 ASR Streaming ([nvidia/nemotron-3.5-asr-streaming-0.6b](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b)) and Nemotron 3 Diarization ([nvidia/Nemotron-3-Diarization](https://huggingface.co/nvidia/Nemotron-3-Diarization)): OpenMDW License 1.1, full text in [`LICENSES/OpenMDW-1.1.txt`](LICENSES/OpenMDW-1.1.txt).
- OpenAI Whisper large-v3-turbo: MIT License ([model card](https://huggingface.co/openai/whisper-large-v3-turbo)).
- UVDoc: Apache License 2.0.
- FrequencyWords word lists: CC BY-SA 4.0, by Hermit Dave.
- EG2 Turkish speech files: CC BY-SA 4.0, see the section below.
These files are modified forms of the original models (format conversion, quantization, repackaging); they are
provided as is, without warranty.
## EG2 Turkish speech model (`eg2_tr_fc8.tflite`, `eg2_tr_3gram_spoken.bin`, `eg2_tr_unigrams.txt`)
### Source
These files are an adaptation of **embeddinggemma2-turkish-stt** by **mstkyvz**
(https://huggingface.co/mstkyvz/embeddinggemma2-turkish-stt), a Turkish speech-to-text model built on
**google/embeddinggemma-2** (https://huggingface.co/google/embeddinggemma-2).
### Licence
- `eg2_tr_fc8.tflite`, `eg2_tr_3gram_spoken.bin`, `eg2_tr_unigrams.txt`: **CC BY-SA 4.0**
(https://creativecommons.org/licenses/by-sa/4.0/), the same licence as the original weights and language model.
Anyone who shares an adaptation of these files must keep the same licence and give the same credit.
- The code of the original project is Apache 2.0. Penra's app code is not part of these files.
- Base model: google/embeddinggemma-2 is licensed Apache 2.0 (https://ai.google.dev/gemma/docs/gemma_4_license, as linked on its
Hugging Face card, checked 2026-11-16). Its weights are not changed and are not part of these files.
- Why share-alike: the original author states the v4/v5 weights and language model are CC BY-SA 4.0 because part of the
training data is share-alike (Wikipedia for the language model, Khan Academy audio). Other data credited by the author:
YODAS (CC BY 3.0), ISSAI Turkish Speech Corpus (MIT), Common Voice (CC0), FLEURS (CC BY 4.0), FineWeb-2 (ODC-By 1.0).
### Changes made by Penra
- Converted the model to LiteRT (`.tflite`) for on-device inference.
- Quantized to int8 ("fc8" quantization of the fully connected layers).
- Baked the prompt-token embeddings into the file.
- Converted the spoken-Turkish 3-gram language model (`lm_spoken/tr_3gram.arpa`) to a compact binary (`eg2_tr_3gram_spoken.bin`); `eg2_tr_unigrams.txt` is the original `lm_spoken/unigrams.txt`.
- No other change to the training data or the training itself. Attribution of the original author does not imply that
mstkyvz or Google endorses Penra or this conversion.
### Use
Used by the Penra app (Pro) for offline Turkish speech to text, downloaded once and run only on the device.
Checksums are in `manifest.json`.