Daili — Model Licenses and Notices

Last updated: October 11, 2026

Daili bundles or optionally downloads the following third-party machine-learning model files. They run locally on the user's device. These notices apply to the model artifacts; they do not change the license of Daili as a whole.

PP-OCRv6 small and PP-OCRv5 Korean mobile

Daili's receipt recognition uses PP-OCRv6 small detection and recognition artifacts derived from PaddleOCR. Devices set to Korean, or located in Korea, instead download the PP-OCRv5 Korean mobile recognition artifact from the same project.

Copyright © 2016 PaddlePaddle Authors. All Rights Reserved.
License: Apache License 2.0
For receipts in Latin-script languages, Daili changed the PP-OCRv6 recognition artifact by slicing its output classes to the Latin subset.
For Japanese receipts, Daili changed the PP-OCRv6 recognition artifact in three ways. Its weights were fine-tuned on text lines of receipt photographs from the datasets credited below. An appended operation limits its output classes to 6,908 of its characters, the ones a Japanese receipt prints, and its character dictionary is cut down to match. Two more appended operations report the most probable character and its probability at each step.
For Chinese and Korean receipts, Daili changed the PP-OCRv6 and PP-OCRv5 Korean recognition artifacts by appending two operations that report the most probable character and its probability at each step. Their weights and character dictionaries are unchanged.
The detection artifact is served without modification.

Read the complete PaddleOCR license

Receipt datasets behind the Japanese recognition weights

The fine-tuning described above used text lines cropped from the receipt photographs of three public datasets. Daili distributes the tuned weights, not the datasets.

JaWildText (Maeda and Okazaki, 2026), Apache License 2.0: the transcribed text lines of its Japanese receipts, which the weights were tuned to read.
CORD v2 (Park et al., 2019), licensed under CC BY 4.0: text lines of its receipts, used without their transcriptions so that the tuned weights go on reading Latin print.
ICDAR 2019 SROIE (Huang et al., 2019), licensed under CC BY 4.0: text lines of its receipts, used the same way.

Moonshine v2 streaming-tiny and streaming-small (English)

Daili's optional voice entry uses the English Moonshine v2 streaming-tiny model as its standard tier and offers streaming-small as an opt-in accuracy tier. Both come from Moonshine AI.

Copyright © 2025 Useful Sensors, Inc. (dba Moonshine AI)
License for the English-language model: MIT License
For the standard tier, Daili converted the original floating-point ONNX graphs to ONNX Runtime format for local mobile inference, quantizing four of the five graphs; the audio frontend is kept at full floating-point precision for accuracy.
The accuracy tier derives from Moonshine Voice v0.1.5's published small-streaming-en/quantized_26_08_21 model, whose audio frontend is stored as 8-bit codes and expanded to floating point on the device when the model loads. Daili stores the tied decoder embedding once instead of serializing its exact transpose twice and retains the adapter positions reachable under Daili's bounded listening duration. Daili calibrates selected weight matrices to 64-value integer palettes and retains 50 sensitive matrices at their original precision. The frontend, embeddings, tokenizer, scales, and inference operators are retained. The resulting model is delivered in reversible DMP1 archives and expanded locally. Calibration changes weights; archive decoding restores those calibrated weights exactly.

Read the complete Moonshine license

FiniteStateEntropy

Daili's model archive decoder includes BSD 2-Clause licensed code by Yann Collet from FiniteStateEntropy. The notice is also included in the Android package's assets. Read the complete license.

Silero VAD v6.2.1

Daili's voice endpoint detector uses the 16 kHz opset-15 ONNX artifact from Silero VAD. It analyzes a detector-only copy of microphone frames; it does not replace or modify the audio sent to Moonshine.

License: MIT License
Upstream version: v6.2.1
Upstream commit: 7e30209a3e901f9842f81b225f3e93d8199902b1
Model SHA-256: 7ed98ddbad84ccac4cd0aeb3099049280713df825c610a8ed34543318f1b2c49 (1,289,603 bytes)

Read the complete Silero VAD license

Return to Daili's privacy policy