Voice technology adoption is on the rise. Gartner forecasts that 2023 is the year when 80% of consumer apps will become “voice-first.” Consumers are using voice-enabled virtual assistants for everything from turning on lights to checking the weather. Relatively recent features like voice-driven online search as implemented by Google are also mainstreaming the use of voice as a means of communication with electronic devices as opposed to the traditional text-based approach.
Businesses are adopting voice recognition for conversational AI to reduce customer support costs and detect intent and sentiment. While AI does not always fully replace human support agents, having an automated system to understand and route customer queries automatically can help to drastically cut down on response and resolution times. Voice biometrics allows businesses to add a new layer of security to verify and authenticate customers’ identities. Media organizations increasingly use AI for automated media captioning.
The AI powering all of these technologies starts with annotated audio data. Audio annotation, a subset of data labeling, is the foundation for building the models used to analyze spoken words, speed up customer responses, or recognize spoken human emotions.
It is essential to differentiate between audio transcription and audio annotation.
Audio or speech transcription is a subset of audio annotation, and it’s the process of converting spoken language into written form. Transcription can be verbatim or non-verbatim. Verbatim transcription includes all conversation filler words, false starts, truncated words or sentences, and pauses. It is a literal conversion of the conversation into written language. Non-verbatim transcription does not include filler words, false starts, truncated words or sentences, and pauses. Transcription provides an easy-to-read version of what has been said in the audio file.
Audio or speech annotation can refer to both the audio transcription and the resulting text’s annotation. An audio annotation is any type of additional information or metadata added to an already existing text. Annotations add phonological, morphological, syntactic, semantic, and discourse information.
Traditionally audio annotation has been a labor-intensive manual task, but in some cases, data scientists can leverage machine learning and natural language processing to automate the process. Some companies, such as Sigma, offer audio annotation as a service, which can help companies leverage the benefits of AI and automated audio annotation without requiring them to develop the infrastructure internally.
Audio annotation services can vary depending on the specific needs of a project. Below are some of the more common types of audio or speech annotation.
Speech-to-text is the transcription from audio to written text. This is useful for many different purposes and is widely used, such as automatically adding subtitles to videos. This can then be further augmented, by offering automated translations of the transcribed text, such as what you’d see on YouTube.
Speech labeling or speech data annotation is when human annotators categorize and label audio data to make it usable for machine learning applications. Keyword or entity annotation involves the extraction of specific keywords or entities such as proper names, telephone numbers, etc. so that AI algorithms can learn how to recognize them or anticipate their usage.
Audio classification involves training models to distinguish particular features within audio samples. These features can cover virtually anything, from distinguishing background noise from relevant speech to identifying what language is being spoken or how many people are talking in a particular sound clip.
Sentiment topic analysis can reveal how a particular person feels about a topic based on the way they talk about it. Sentiment analysis done over audio is generally considered more accurate than text-based sentiment analysis, as there are more cues that can be used to gauge emotion compared to text. Not only the content of the speech itself is considered but also things like voice volume, tone, and cadence can all be used as part of sentiment analysis.
This involves judging the intent of a user based on what they have said. Conversational AI must deal with a wide variety of different phrasings or ways to communicate a particular intention. Casual intent deals with more generic scenarios – an example of this would be a user saying either “hello”, “hey”, or “hi”, which would generally prompt a bot to respond in kind with a greeting as the intent is to start a conversation. The other type of intent is business intent, which deals more specifically with information pertinent to the specific business interests of an organization or its clientele.
Virtual assistants like Amazon’s Alexa or Apple’s Siri use audio annotation to understand and respond to user queries. The speech recognition software used by these virtual assistants must be trained with a large amount of data to understand the user accurately. This data is typically annotated so the machine learning algorithms can learn from it.
In-vehicle navigation systems also use audio annotation to provide voice guidance to drivers. The system must be able to understand the driver’s speech to provide accurate directions.
Call center responses are another area where audio annotation is used to provide accurate responses to customer queries, the call center agents must be able to understand the customer’s speech. This data is typically annotated so the machine learning algorithms can learn from it.
Implementing audio annotation in a project can be a daunting prospect, particularly for organizations that do not have experience in the field. That’s why many organizations choose to contract outside help and many choose to use Sigma.ai to help guide them by following the Sigma process.
- Project Analysis: The Sigma process begins by devoting some of our senior leaders to analyze the project, discuss details and develop a project proposal.
- Project Manager Assignment: From there, a project manager will be assigned to the project based on their skills and expertise as it relates to the project’s needs. A project advisory committee (PAC) is formed with internal experts in the project area who then discuss the project in detail with the client to disambiguate any issues and review the project requirements.
- Guidelines & Requirements: Discuss the nuances of annotation guidelines with the client and review projects requirements as well as address objectives, metrics, deadlines, deliverable and security and privacy needs.
- Tools & Procedures Setup: Once details are confirmed, the tools and procedures are set up, built, and optimized for the particular project to maximize throughput and quality. The project definition is then refined.
- Comprehensive Test: Annotators use the adapted tools and procedures to annotate or collect sample data, including for quality assurance (QA). The PAC produces a report outlining the results and any suggestions. This report and sample data are sent to the client for review and for any applicable retesting or retooling based on client feedback.
- Client Feedback: The tools and processes are optimized following client’s comments. If the client has updated the guidelines, the tools are retested.
- Annotation Collection at Scale: Experienced annotators are then brought on as part of an iterative process to produce data at scale and improve the performance of the AI.
- Quality Assessment: A final quality assessment report is produced and delivered to the client at the completion of the project.
Because Sigma assigns a specific team of experts built for each individual project, it has been shown that their expertise can help increase productivity by 25% and project quality by 32% on average.
Audio annotation is widely used in artificial intelligence projects, particularly those that involve interacting with consumers, such as call center automation, voice biometric security systems, and virtual assistants. Conversational AI allows consumers to ‘talk’ to devices in much the same way as they would to another human being. Audio annotation is particularly important in the modern-day due to the prevalence of devices like cell phones and tablets with microphones and the superior user experience offered by enabling speech interaction for virtually everything.
Growth in AI markets dealing with virtual assistants continues unabated, generating $3.9 billion of revenue for enterprises in 2020 with a compound annual growth rate of 28% predicted from 2022-2032.
Audio annotation is used across several distinct industries for a variety of purposes. Audio annotation is an important part of developing virtual assistants and conversational AI that respond to human speech and voice commands. As a direct example, virtual assistants like Amazon Alexa and Apple’s Siri need to know when they are being spoken to and issued a command as opposed to when they are simply picking up background voices in the environment that aren’t directed at them.
There are several types of audio annotation and each has distinct purposes and use cases. Some of the most popular types of audio annotation include the following:
Transcription from speech to text is both a middle and end step for many different types of audio annotation. Sometimes transcription is all that is needed for a task, but sometimes transcription simply represents one stage in a larger process where transcribed speech is then used for other natural language processing tasks. Speech-to-text is often used in call center applications to make sense of the voices of their customers.
Utterances, intents & entities are three categories that are commonly used in natural language processing to understand and process a command. An utterance is simply what is being said. A voice command, such as “show me the weather for tomorrow” is an utterance.
Intent is the meaning behind those words. There is a lot of ambiguity in language. Humans tend to deal with this naturally, using context and other clues to resolve ambiguities, often quickly and subconsciously . However, machines need to clearly understand the intent behind an utterance. Intent is usually defined using verbs and nouns. For example, in the command, “show me the weather for tomorrow”, the intent could be defined as ‘show weather’. This is a verb and noun that concisely defines the intent.
Entities are additional modifiers that can aid or alter the intent. In our example given above, ‘tomorrow’ is an entity. It specifies a time that does not change the overall intent but acts as a modifier as to which weather results to return. Location is another example of an entity, though in this case, one is not specified, so other contextual cues must be used to determine what location to pull weather forecasts from.
Speaker classification involves determining whether a particular audio sample is human speech or not and identifying the properties of the voice used to produce sounds. This allows for multiple different voices to be individually distinguished in audio samples. This is useful for virtual assistants, who can be trained to only respond to particular voices rather than anyone who issues a comprehensible command.
Music classification is a special type of audio classification. Its purpose is to pick up on specific aspects of a musical audio file. This can be used to identify the genre of a track or to identify and isolate instruments and vocals within a song. A common use for this type of technology is for the automatic categorization of music and building playlist recommendation engines based on personalized music preferences.
Good data preparation practices can make life much easier for audio annotation. Data preparation is a multi-step process that optimizes data for use with machine learning. In particular, data used for machine learning model training should be prepared with best practices in mind as the training data input is crucial to maximizing the success of the resulting model and its ability to generalize data that it has not been trained for.
Good data preparation practices as they relate to audio annotation include good data sourcing, cleansing and annotation. Data sourcing refers to the practice of sourcing data for machine learning. Machine learning requires high volumes of high-quality data, which is not always easy to source. That’s why data cleansing is also important – this involves cleaning up data, filtering out anything unnecessary or irrelevant and standardizing it such that all data follow a consistent format. Data annotation or labeling can also help as it enriches data, providing metadata that can help with the categorization, sorting, and storing of data.
Manual annotation is the process of having humans listen to and annotate audio data. This is a time-consuming process, but it is important for the creation of high-quality training datasets. There are several ways in which organizations can arrange for audio to be manually annotated, though each has its own strengths and weaknesses. An organization may choose to deal with audio annotation tasks in-house, outsource them, or crowdsource them.
Dealing with audio annotation tasks in-house means that you have full control over the whole process and can control for quality but it is costly. Outsourcing the work is slightly cheaper but there is less control over the quality and there may be some concerns around data privacy regulations if personal or sensitive data is processed. Crowdsourcing is the cheapest of all but also typically the lowest in quality and is unsuitable for dealing with data of a private or sensitive nature.
The main advantages of manual annotation are that is is more accurate and can be used to annotate a wide range of different types of data. However, the drawbacks compared to AI-powered annotation are that they are expensive, slow, and could be of variable quality. You also have to concern yourself with the rules and regulations of handling private data unless the company you outsource annotation work to, like Sigma.ai, has comprehensive security measures or secure facilities to deal with data security and privacy.
Audio annotation is fast becoming an important part of customer-oriented business practices.
Virtual assistants and other tools that rely on audio annotation can understand and respond to your customers’ requests even when they are phrased in a very natural and conversational manner. This means customers can simply describe their problem rather than having to follow specific procedures or press particular buttons. However, a lot of training data is needed, particularly for domain-specific knowledge, so that these machine learning models are capable of handling a wide variety of tasks related to responding to customers.
A study of the barriers to audio or voice technology annotation includes accuracy (73%), accent or dialect recognition issues (66%) and language coverage (38%).
Sigma guarantees 98% accuracy using our QA methodology and tools, and up to 99.99% upon request. We support over 250+ languages and dialects and rely on over 22,000+ highly trained annotators, linguists and subject matter experts in 5 continents, with access to a pool of over 900,000.
If you are looking to implement audio annotation as part of a project, contact us. Our team can help to guide you through the process so that you get the best results for your next audio annotation project.