⚒️Use-cases

The first version of Ta-da will offer two types of jobs. This section describes them and lists additional ones we will add in further versions of Ta-da.

Scripted Audio Recordings

Scripted audio recordings are datasets of audio samples that can be used to train and test speech recognition models. These recordings help machine learning models to learn how to recognize different accents and dialects, as well as how to identify different words and phrases. They can also be used to create datasets that are specific to certain domains, such as medical speech or customer service conversations.

The main method for making scripted audio recordings for machine learning is to create a script and then record it. The script should include information about the intended audience, the content that needs to be recorded, and any relevant audio cues.

These data are particularly difficult to obtain because they need to be extremely varied. Indeed, for a voice assistant to recognize any type of voice, it must be trained with recordings from men and women of all ages, sometimes with background noise, in different ways of speaking (whispering, shouting, etc.), with specific vocabulary, etc. Thanks to the ease of use of Ta-da, anyone can record their voice and participate in creating diverse datasets, which is an essential need to create good datasets.

Image labeling

Image labeling is the process of assigning labels to an image or set of images. Labels can be as simple as classifying an object as a “cat” or “bicycle”, or as complex as recognizing an action in a sequence of images. Image labeling is an important component of many computer vision applications, including object detection, scene understanding, and image classification. Here are some example you may know:

• The auto-pilot of Tesla uses these datasets to train on recognizing the road, the pedestrians, the signs, etc.

• The facial recognition systems on your smartphones to unlock it are also trained on these types of datasets.

There are several types of image annotation. The first version of Ta-da offers bounding boxes. A bounding box is a rectangular box that is drawn around an object in an image. It defines the area of the object and is used to label the object by specifying the coordinates of the box. For example, a bounding box for an image of a cat might be drawn around the cat. Technically speaking, we store the coordinates of the box’s top left corner, the width and height of the box, and the class of the object (e.g. cat). Below is an example of a bounding boxe draw around a cat:

Upcoming features

We know that machine learning algorithms are diverse and need a wide variety of data. This is why we plan to implement other types of jobs. This section gives a brief overview of what you can expect from us in further versions.

Voice data

Let's start with the audio data. As explained earlier, Ta-da offers scripted audio recording jobs in its first version. This means that a producer records himself while he reads a sentence. This type of job is interesting because it allows to easily collect audio data with a fixed vocabulary and strict rules. However, in order to get more efficient voice assistants, another category of data is required: spontaneous data. Indeed, when you talk to another person, the way you formulate sentences is different than when you read. This difference is crucial and has an influence on the quality of voice assistants. Based on this knowledge, we will soon implement a type of job that allows us to collect spontaneous data.

Natural language processing data

Natural Language Processing is also a huge field of research, especially Natural Language Understanding (NLU). In order to support this industry, we will develop several types of jobs such as:

• Translation: the data collected from a translation job is used to train machine translation AI for automatic translation solutions. You have probably used online translation solutions before.

• Text classification: the process of assigning a label or class to a piece of text. This label may indicate the type of content, such as news, opinion, reviews, etc. The criteria may include keywords, the length of the text, the number of words, or other features of the text. Text classification can be used to help organize large amounts of data and to identify trends in the data.

• Token classification: a type of text classification that assigns labels to individual words or tokens in a text. This type of classification can be used to identify the parts of speech in a sentence or to determine the sentiment of a sentence. For example, this is useful for companies that want to know what reputation they have on social media.

• Summarization: condensing information into a concise and comprehensive summary. It is a way of quickly extracting the most important and essential information from a text, while still maintaining its key points and themes. Summarization can be used to train AI to simplify texts, to extract key points, or to provide an overview of the content.

Content managment

In the dynamic sector of retail and e-commerce, effective Product Information Management (PIM) is crucial for business success. A key component of this management is the use of structured data, where each product is meticulously documented with well-defined labels. This structuring not only allows for better organization of product data but also significantly optimizes user experience. For instance, precise labels facilitate faster and more accurate product searches, thus improving the online shopping experience. Moreover, in an environment where artificial intelligence and machine learning are increasingly influential, well-structured data is indispensable for feeding personalized and relevant product recommendations. Therefore, the implementation of robust and intelligent PIM systems, capable of efficiently managing these structured data, is a major challenge for businesses looking to remain competitive in the digital commerce landscape.

To further enhance the efficacy of the PIM systems in the retail and e-commerce sector, we will introduce a range of new job types within our platform. These jobs are specifically designed to engage our user base in the detailed process of product data labeling and verification. This approach is particularly effective in complex categories where nuanced understanding of products is essential. The active participation of users in these new roles will be instrumental in refining the data quality, thus making PIM systems more robust, intelligent, and tailored to the dynamic demands of the digital marketplaces.

Image and video data

In the field of Computer Vision, there are seven methods of labeling. We have selected bounding boxes as the first feature. Nevertheless, we plan to implement many others:

• Classification: the process of assigning a label or class to an image, such as a person, object, scene, or activity.

• Polygons: a type of image annotation used to label the boundaries of objects in an image. It involves manually drawing polygons around objects of interest in an image, such as cars, buildings, people, etc. This type of annotation is commonly used to train deep learning algorithms for object detection and segmentation.

• Semantic segmentation: used to assign a semantic label to each pixel in an image, such as "cat" or "road." These labels can then be used to classify the image into different objects and scenes. This technology is useful in a variety of applications, such as autonomous driving, medical imaging, and satellite image analysis.