What algorithm does google tesseract-OCR use for recognition?

This article in tesseract source provides an in-depth overview of the technology.

In particular:

Blobs are organized into text lines, and lines and regions are parsed into fixed pitch or proportional text.

[...]

Recognition then continues as a two-pass process. The first pass tries to recognize each word in turn. Each word that is satisfactory is passed on to the adaptive classifier as training data. the responsive classifier is then able to more accurately recognize the text below the page.

[...]

After the text lines have been found, the baselines are set more accurately using a quadratic spline.

[...]

Baselines are established by dividing blocks into groups using a reasonable continuous offset for the original straight baseline. A quadratic spline is set to the most populous section (presumably as a baseline) using the least squares method.

The doc doesn't explicitly state if it uses a neural network, but given the content, I would argue that it is possible, at least for parts of it.

For more information on finding a line, see R. Smith, “A Simple and Efficient Skew Detection Algorithm Using Accumulated Text,” Proc. 3rd International. Conf. on Document Analysis and Recognition (Volume 2), IEEE 1995, pp. 1145-1148.