HMMs vs RNNs for Speech Recognition

Comparison of Hidden Markov Models and Recurrent Neural Networks for Speech Recognition

• Focus on comparing Hidden Markov Models (HMMs) and Recurrent Neural Networks (RNNs) for speech recognition, highlighting their strengths and weaknesses.
• Explores challenges in employing HMMs and RNNs for speech recognition like the need for extensive training data and managing speech variability.
• Touches on future research directions, including developing more robust, accurate models and exploring new neural network architectures.

Introduction

• Speech recognition is a field aimed at enabling computers to understand and transcribe human speech, with applications including virtual assistants, voice dictation, and voice control systems.
• Various approaches include pattern matching, dynamic programming, and statistical methods.
• HMMs have been used for speech recognition, modeling speech as a sequence of hidden states corresponding to phonemes, with parameters estimated from training data using the Expectation-Maximization (EM) algorithm.
• RNNs have emerged as a popular alternative, learning temporal dependencies to map acoustic features to phoneme sequences and showing promise, especially combined with deep learning techniques.
• A comparison between HMMs and RNNs for speech recognition is made, highlighting strengths, weaknesses, challenges, and future research.

Hidden Markov Models and Recurrent Neural Networks

Hidden Markov Models

• HMMs are statistical models assuming the system being modeled is a Markov process with unknown states, characterized by a set of states, transition probabilities between states, and emission probabilities for each state.
• Transition probabilities specify the likelihood of moving from one state to another, and emission probabilities define the likelihood of emitting a certain observation given a state.
• For speech recognition, HMM states typically correspond to phonemes or phonetic sub-units, while observations are acoustic features extracted from the speech signal, such as Mel-Frequency Cepstral Coefficients (MFCC).
• Training an HMM involves estimating transition and emission probabilities from training data, commonly using the Expectation-Maximization (EM) algorithm, which iteratively alternates between the Expectation (E) and Maximization (M) steps.
• The E step calculates the probability of being in each state at a given time, given the observations and current parameter estimates, and the M step updates parameter estimates to maximize the probability of the observations, given the state probabilities.
• After training, an HMM can recognize new speech by finding the most probable state sequence that generates the observed acoustic features.