How Speech Recognition Works to Enhance our Daily Lives
Speech and voice recognition technology is integrated into many parts of our lives. Whether it’s the voice assistant on your phone, the virtual assistant in your car, or a customer service chatbot from your favorite online store, chances are high that you interact with speech recognition technology on a regular basis.
Besides assisting consumers to check the weather, message friends, and set reminders, speech and voice recognition technology is used by a variety of industries to streamline and improve processes: the healthcare industry, for example, uses it to help with diagnostic reporting, virtual consultations, and writing administrative reports. The financial industry applies it to provide research and analytics reports to their customers, while call centers use it to answer customer queries quickly and efficiently.
So profound is the impact of speech technology on our lives, that many experts believe it will evolve to become the default way humans interact with machines.
“The most important thing to remember is that speech is the most natural form of human communication. Writing systems are arbitrary and have to be learned, whereas learning to speak is a natural process: you learn by listening to other speakers around you. The question then becomes: why type when you can talk? This is why speech technology is the future,” said Christopher Shulby, Director of Machine Learning at DefinedCrowd.
In this blog post, we’ll be taking a closer look at speech recognition technology: what it is, how it works, and how training datasets enable it to understand human speech in increasingly complex ways.
Speech Recognition Basics: What It Is and How It Works
Speech recognition refers to technology that allows machines, using a variety of systems, programming and algorithms, to recognize and make sense of human speech and convert it into text. From the text form, it is able to extract meaning and understand what is being said. It is also known as Automatic Speech Recognition (ASR).
It works, basically, by taking recordings of human speech gathered into datasets, and breaking these down into smaller and smaller bits of information – from full speech to utterances and their corresponding transcriptions (text). From these audio samples and text transcriptions, the technology learns to recognize and interpret more complex speech patterns, vocabulary, and meaning.
Understanding speech, however, is more complicated than that. Machines also need to understand utterances spoken in a variety of settings, for example when people are speaking in noisy environments (from the metro or bus, or around building works), or if a person is speaking from afar or at a low volume.
Additionally, the technology must take into account variations in accent, syntax, local expressions, and different ways of saying the same things within each individual language.
The future of speech technology will be focused on developing a more robust and accurate response to adverse conditions, such as background noise, distance from the microphone, and other factors that currently affect the quality of many speech recognition systems.
There are two different types of speech recognition: speaker-dependent and speaker-independent.
Speaker-Dependent Speech Recognition
Speaker-dependent speech recognition is a type of speech recognition that is based on one voice, and trained to that one person’s specific accent, expressions and way of speaking. According to Shulby, speaker-dependent speech recognition’s main aim is to improve accuracy in speech recognition:
“The biggest hindrance to the adoption of speech technology is the accuracy itself. If we had close to 100% accuracy in speech recognition, we would use it across the board. However, most broad speech recognition tasks have a 30% error rate. Speaker-dependent technology tries to solve that. Focusing on one speaker boosts the accuracy, as it is adapted to your voice.”
With speaker-dependent speech recognition, variations of speech such vowel pronunciation, accents and dialects are removed so that the system is trained only on one specific voice and its variations.
The obvious downside, then, is that speaker-dependent speech recognition can only be used effectively by the one person the system was trained on, which requires a small amount of training to reduce errors and improve recognition.
Fields of application:
Speaker-dependent speech recognition is most often used for dictation software, for use in industries such as healthcare, where doctors can take accurate voice notes (which are converted into text) of case history and diagnostic situations.
Speaker-Independent Speech Recognition
Conversely, speaker-independent speech recognition is developed to respond to anyone’s voice, and is trained to recognize utterances in the variety of ways in which they can be said. With this type of speech recognition, some accuracy is sacrificed due to all the variations in speech, but as an advantage, it can be effectively used by a much broader group of people and does not require training before application in any situation.
Fields of application:
Speaker-independent speech recognition then has much broader application possibilities than its counterpart. It is used, for example, in call centers, where an Integrated Voice Response (IVR) system allows the user to use speech to navigate menus.
Speaker-independent is also used in virtual assistants, in-car navigation and other situations in which the machine needs to understand a standard set of commands (“turn up the music, what is the weather tomorrow, where is the nearest pharmacy”).
In order for speech recognition technology to work accurately, it’s necessary to train them on existing data to help them recognize words and sounds, and, eventually, extract meaning and intent behind those words. There are several types of training data that can be used for training purposes.
Scripted Speech Data
Scripted speech data consists of data collected from recordings of a variety of people reading from the same script, creating a broad set of utterances of the exact same script, with all its variations.
This type of speech data is most often used to train systems that need to react to voice commands and specific words. An example of this would be in-home assistants such as Alexa or Google Home, who need to understand simple commands such as: “Turn off the lights.”
Conversational Speech Data
This type of speech data is more complex than scripted speech, as it is training the system to interpret spontaneous dialog and recognize context, rather than simply the specific words. Conversational speech data uses recordings of conversations between two people and aims to train systems to recognize when the subject of the conversation has shifted abruptly, or when understanding of the words relies heavily on context.
Shulby explained further. “If you are speaking about a music festival, or having a serious conversation about WW2, the tone, context and vocabulary is very different. Conversational speech datasets then help to give context for better accuracy.”
This type of data is useful for highly customized use cases such as chatbots. As customers begin to expect more complex reactions and information from chatbots, training with conversational speech data will help improve accuracy and understanding of human speech.
The Difficulty of Getting Data
Data scientists and machine learning engineers realize how crucial high-quality datasets are to the success of the speech recognition model. Datasets need to be large, diverse (representing all accents and dialects), and high-quality. However, getting ahold of these datasets is easier said than done. In fact, many studies cite access to data as one of the main barriers to the widespread adoption of artificial intelligence (AI). And data scientists often just don’t have the time to collect, annotate and validate their own AI training datasets, which can take months.
Put simply, the world needs more readily available, off-the-shelf datasets to encourage the growth of AI. DefinedCrowd is on it!
Free Speech Recognition Dataset Samples
Access to accurate and comprehensive datasets is essential to successful use of speech recognition technology in a variety of use cases and to speed up time to market.
DefinedData was created to meet this need. DefinedData is an online catalog of off-the-shelf datasets, available for immediate use. Datasets within this catalog are pre-collected, annotated and validated by a global crowd of over 450, 000 contributors, ensuring the high-quality DefinedCrowd has become known for. These datasets can be used to effectively train baseline models, improve and expand existing models, or benchmark third-party applications.
These are just a few of the speech recognition datasets available with DefinedData. For each dataset, prospective customers are able to request a free sample to test the quality for themselves.
For a full list of datasets that DefinedData has to offer, see our online catalog.
Speech recognition is an exciting aspect of AI and machine learning with almost limitless applications. It allows humans to be understood by machines in increasingly advanced ways. With the right training data, ASRs can be tailored to one specific person, or can be trained to be used effectively by anyone. Read more here about how DefinedData can help improve ASR performance and results.