A local AI for Gramps?

,

Other engine (toolset)

On the handwritten text, it is still not usable (OCR).

{
  "execution_time": "PT2.979285S",
  "results": [
    {
      "text": "NOMS",
      "confidence": 0.943,
      "polygon": [
        [
          453,
          20
        ],
        [
          481,
          19
        ],
        [
          481,
          28
        ],
        [
          453,
          28
        ]
      ]
    },
    {
      "text": "FOLIOS",
      "confidence": 0.758,
      "polygon": [
        [
          552,
          14
        ],
        [
          585,
          13
        ],
        [
          586,
          22
        ],
        [
          553,
          24
        ]
      ]
    },
   ...,
    {
      "text": "PRENOMS ET DEMEURES",
      "confidence": 0.626,
      "polygon": [
        [
          35,
          28
        ],
        [
          125,
          29
        ],
        [
          125,
          39
        ],
        [
          35,
          37
        ]
      ]
    },
    {
      "text": "FALNOMS ET DEMELHES",
      "confidence": 0.6946666666666667,
      "polygon": [
        [
          230,
          32
        ],
        [
          318,
          33
        ],
        [
          318,
          42
        ],
        [
          230,
          42
        ]
      ]
    },
    {
      "text": "FARNOMS KT DEMEURES",
      "confidence": 0.4036666666666667,
      "polygon": [
        [
          420,
          33
        ],
        [
          515,
          29
        ],
        [
          516,
          38
        ],
        [
          420,
          42
        ]
      ]
    },
    {
      "text": "MATRICE",
      "confidence": 0.568,
      "polygon": [
        [
          552,
          41
        ],
        [
          583,
          40
        ],
        [
          583,
          48
        ],
        [
          553,
          49
        ]
      ]
    },
    {
      "text": "HATAICE",
      "confidence": 0.607,
      "polygon": [
        [
          354,
          47
        ],
        [
          384,
          47
        ],
        [
          384,
          54
        ],
        [
          354,
          55
        ]
      ]
    },
    {
      "text": "LE PROPRIETAIKIA.",
      "confidence": 0.1705,
      "polygon": [
        [
          434,
          47
        ],
        [
          500,
          44
        ],
        [
          500,
          52
        ],
        [
          434,
          54
        ]
      ]
    },
    {
      "text": "Artegala yetin",
      "confidence": 0.129,
      "polygon": [
        [
          202,
          82
        ],
        [
          352,
          82
        ],
        [
          352,
          104
        ],
        [
          202,
          106
        ]
      ]
    },
    {
      "text": "Arbogail Joual Autorfurent",
      "confidence": 0.15866666666666665,
      "polygon": [
        [
          204,
          111
        ],
        [
          351,
          116
        ],
        [
          350,
          130
        ],
        [
          203,
          127
        ]
      ]
    },
   ...

OK, my example is not in a “lab” environment (resolution, drop shadow, strikethrough text, degraded format, multiple scripts and authors, etc.), yet there are still nearly 50% errors on the printed French text.

On the handwritten text (text and numbers), it’s closer to 99% errors… Most humans capable of reading this kind of text could achieve better results!