cstr/tesseract-lstm-GGUF

Tesseract LSTM OCR — GGUF models for CrispEmbed

Tesseract's LSTM line-recognition engine converted to GGUF format for CrispEmbed inference. Supports 12 languages with models ranging from 250 KB to 3.2 MB.

These are the neural network weights from tessdata_best (float64 source), converted with full precision then quantized. The LSTM recognizes pre-cropped text line images — pair with a text detector (DBNet, Surya) for full-page OCR.

Models

Language	Code	Params	F16	Q8_0	Q4_K	VGSL topology
English	eng	1.46M	2.9 MB	1.5 MB	857 KB	Lfys64 Lfx96 Lrx96 Lfx512
German	deu	940K	1.8 MB	976 KB	537 KB	Lfys64 Lfx96 Lrx96 Lfx384
French	fra	391K	769 KB	435 KB	270 KB	Lfys48 Lfx96 Lrx96 Lfx192
Spanish	spa	1.46M	2.9 MB	1.5 MB	853 KB	Lfys64 Lfx96 Lrx96 Lfx512
Italian	ita	822K	1.6 MB	860 KB	481 KB	Lfys64 Lfx64 Lrx64 Lfx384
Portuguese	por	822K	1.6 MB	860 KB	481 KB	Lfys64 Lfx64 Lrx64 Lfx384
Dutch	nld	408K	801 KB	449 KB	247 KB	Lfys64 Lfx96 Lrx96 Lfx192
Russian	rus	1.47M	2.9 MB	1.5 MB	860 KB	Lfys64 Lfx96 Lrx96 Lfx512
Arabic	ara	1.45M	2.8 MB	1.5 MB	848 KB	Lfys64 Lfx96 Lrx96 Lfx512
Chinese (Simplified)	chi_sim	1.52M	3.1 MB	1.6 MB	900 KB	Lfys64 Lfx96 Lrx96 Lfx512
Japanese	jpn	1.62M	3.2 MB	1.7 MB	952 KB	Lfys64 Lfx96 Lrx96 Lfx512
Korean	kor	1.48M	2.9 MB	1.5 MB	863 KB	Lfys64 Lfx96 Lrx96 Lfx512

Quantization quality

Tested with crispembed-diff per-layer parity harness against F16 Python reference:

Quantization	Cosine similarity (worst layer)	Text output	Recommendation
F16	1.000000	Perfect match	Best quality, 2x size of F32
Q8_0	0.9994+	Identical to F16	Recommended — no text degradation
Q4_K	0.92-0.99	Core text correct, minor edge artifacts	Smallest size, acceptable for most uses

Usage

CLI

crispembed -m tesseract-eng-q8_0.gguf --ocr image.png

C API

#include "tesseract_lstm.h"

tesseract_lstm_context *ctx = tesseract_lstm_init("tesseract-eng-q8_0.gguf", 4);
const char *text = tesseract_lstm_recognize(ctx, gray_pixels, width, height, NULL);
printf("%s\n", text);
tesseract_lstm_free(ctx);

Unified OCR API (auto-detects architecture)

void *ctx = crispembed_math_ocr_init("tesseract-eng-q8_0.gguf", 4);
const char *text = crispembed_math_ocr_recognize(ctx, pixels, w, h, channels, NULL);
crispembed_math_ocr_free(ctx);

Architecture

The VGSL (Variable-size Graph Specification Language) network:

Input (grayscale, height-normalized)
  -> Convolve 3x3 stacking (no weights, just neighborhood concat)
  -> FullyConnected + tanh (9 -> 16 features)
  -> MaxPool 3x3
  -> XYTranspose + SummLSTM (summarize height dimension)
  -> LSTM forward (96 or 64 units)
  -> LSTM reverse (96 or 64 units)
  -> LSTM forward (192-512 units)
  -> Softmax output
  -> CTC greedy decode

Spaces are not produced by the LSTM — Tesseract adds them via its DAWG language model (word segmentation), which is not part of this engine. Use the output as character-level recognition and add word boundaries via the text detector's bounding boxes.

Conversion

Convert any of the 126 Tesseract languages from .traineddata:

python models/convert-tesseract-to-gguf.py \
    --model eng.traineddata --output tesseract-eng-f16.gguf --fp16

# Then quantize:
crispembed-quantize tesseract-eng-f16.gguf tesseract-eng-q8_0.gguf q8_0
crispembed-quantize tesseract-eng-f16.gguf tesseract-eng-q4_k.gguf q4_k

Supports both tessdata_best (float64, recommended) and tessdata_fast (int8).

Source

Weights from tesseract-ocr/tessdata_best (Apache-2.0 license). Converted with CrispEmbed's convert-tesseract-to-gguf.py.