Table 1 Explanation of technical terms

Beam search

A heuristic search algorithm. In Chiron, the beam search decoder with beam width W maintains a list of the W most probable sequences up to position i and constructs the probabilities of all possible sequence extensions for i + 1.

[32]

Connectionist Temporal Classification (CTC) decoder

A type of neural network output and scoring for labeling sequence data with RNNs. It does not require presegmented training data and postprocessed outputs.

[56]

Convolutional Neural Network (CNN)

A type of neural network often used for image analysis. It can recognize patterns by applying different filters to an image.

[57]

Forward algorithm

An algorithm that computes the probability P(x) of a sequence x given a certain HMM.

[58]

Hidden Markov Model (HMM)

A stochastic model that models a sequence of unobserved events underlying a sequence of observations. HMMs assume that an event only depends on the previous event.

[58, 59]

Long-short-term memory (LSTM) unit

A type of RNN that can be used as a building block in bigger networks. It has specific input, output, and forgot gates that allow it to retain or discard information that was passed on from a previous state.

[60, 61]

Partial Order Alignment (POA) graph

A graph representation of a multiple alignment that allows each base in the alignment to have multiple predecessors. Different paths through the graph represent different alignments.

[62]

Recurrent Neural Network (RNN)

A type of neural network that takes information passed on from previous states into account.

[63]

Training data

A dataset that is used to optimize (i.e., train) the parameters of a model. Training is required for both HMMs and RNNs. The training dataset thus determines the performance of the model.

[58, 63]

Viterbi decoding

An algorithm that finds the most likely sequence of events given a certain HMM.

[58]