Definitions: Phoneme – Distinct sounds in a language Grapheme – Distinct unit written in a language Lexicon – Dictionary that maps phoneme to graph(i.e. Arpabet) Architecture design choices of the above for ASR : 1. Target audio straight to Grapheme(without lexicon) – so basically extract features to Graphemes(words) – usedContinue Reading
CBOW learns to predict the word by the context window(+nt -nt words) around it by taking the max probability of the word that fits this context(the most frequent word it learns).Therefore words with infrequent values will not work great with this technique. Skip-Gram is inverted to CBOW, thus learns toContinue Reading
During inference, precision in floats is not needed and can be reduced to using 8 bits instead of 32 bits this allows to bin continuous values to discrete ranges, and therefore is known as Quantization. This enables us to increase the bandwith sent (or increase the memory footprint in caseContinue Reading
After training we can optimize a frozen graph or even a dynamic graph by removes training-specific and debug-specific nodes, fusing common operations, and removes code that isn’t used/reached. Code Example from tensorflow.python.tools import optimize_for_inference_lib inputGraph = tf.GraphDef() #read in a frozen model with tf.gfile.Open(‘frozentensorflowModel.pb’, “rb”) as f: data2read = f.read() inputGraph.ParseFromString(data2read) outputGraph = optimize_for_inference_lib.optimize_for_inference(inputGraph, [“inputTensor”], Continue Reading
In prediction/inference mode, variable types are unnecessary, so by freezing the graph we convert all variables in a graph and checkpoint into constants. Also there are structure nodes and operations related to training that we don’t need and can be removed by the freezing process as well. The freezing processContinue Reading
Hidden Markov Model (HMM) is a statistical Markov model in which the system being modeled is assumed to be a Markov process with unobservable (i.e. hidden) states. In simpler Markov models (like a Markov chain), the state is directly visible to the observer, and therefore the state transition probabilities areContinue Reading
Hidden states are the unknowns we try to detect or predict. The Hidden states have a relationship amongst themselves called the transition probabilities. Observations are the evidence variables that we have a priori. Observations and states have a relationship between them called the emission probabilities.Continue Reading
A Markov chain is a stochastic model describing a sequence of possible events in which the probability of each event depends only on the state attained in the previous event. (“memorylessness”)Continue Reading
The Viterbi algorithm (using the maximum likelihood decoding (MLD) algorithm) is a dynamic programming algorithm for finding the most likely sequence of hidden states – called the Viterbi path in hidden Markov models (HMM)Continue Reading
A variational autoencoder provides a probability distribution for describing an observation/attribute in latent/hidden space.Continue Reading