Natural Language Processing

Automatic speech recognition has a lot of potential applications, such as audio transcription, dictation, audio search, and virtual assistants. I am sure that everyone has interacted with at least one of the virtual assistants by now, be it Apple's Siri, Amazon's Alexa, or Google's Assistant. At the core of all these speech recognition systems are a set of statistical models over the different words or sounds in a language. And since speech has a temporal structure, HMMs are the most natural framework to model it.

HMMs are virtually at the core of all speech recognition systems and the core concepts in modeling haven't changed much in a long time. But over time, a lot of sophisticated techniques have been developed to build better systems. In the following sections, we will try to cover the main concepts leading to the development...

Part-of-speech tagging

The first problem that we will look into is known as part-of-speech tagging (POS tagging). According to Wikipedia, POS tagging, also known as grammatical tagging or word-category disambiguation, is the process of marking up a word in a text as corresponding to a particular part of speech based on both its definition and its context, that is, its relationship with adjacent and related words in a phrase, sentence, or paragraph. A simpler version of this, which is usually taught in schools, is classifying words as noun, verbs, adjectives, and so on.

POS tagging is not as easy as it sounds because the same word can take different parts of speech in different contexts. A simple example of this is the word dogs. The word dogs is usually considered a noun, but in the following sentence, it acts like a verb:

The sailor dogs the hatch.

Correct grammatical tagging...

Speech recognition

In the 1950s, Bell Labs was the pioneer in speech recognition. The early designed systems were limited to a single speaker and had a very limited vocabulary. After around 70 years of work, the current speech-recognition systems are able to work with speech from multiple speakers and can recognize thousands of words in multiple languages. A detailed discussion of all the techniques used is beyond the scope of this book as enough work has been done on each technique to have a book on itself.

But the general workflow for a speech-recognition system is to first capture the audio by converting the physical sound into an electrical signal using a microphone. The electrical signal generated by the microphone is analog and needs to be converted to a digital form for storage and processing, for which analog-to-digital converters are used. Once we have the speech in digital...

Summary

In this chapter, we looked into two of the major applications of HMMs: POS tagging and speech recognition. We coded the POS tagger using a most-frequent tag algorithm and used the pomegranate package to build one based on HMM. We compared the performance using both these methods and saw that an HMM-based approach outperforms the most-frequent tag method. Then, we used the SpeechRecognition package to transcribe audio to text using Google's Web Speech API. We looked into using the package with both audio files and live audio from a microphone.

In the next chapter, we will explore more applications of HMMs, specifically in the field of image recognition.