AI Datasets & Machine Learning Services

An AI Dataset is a large amount of data that is typically delivered in the form of audio. It can also be text-based if a company has already transcribed the data; however, we’ll need the audio to cross check the dataset’s accuracy. In terms of size, TransribeMe typically receives and reviews 5,000-10,000+ hours of audio for any given project – but we can start with less and work on it in batches together.

Companies will use TranscribeMe’s AI Dataset service to help train a computer to perform a specific automated function, such as taking orders at a drive-through, directing service requests over the phone, or developing a chatbot. Performing these functions necessitates a large amount of highly accurate annotated data to train the computer on how to interact with humans in order to complete specific tasks.

You can train computers to recognize what certain words mean and what else is associated with those words. For example, if you have fries, fries and gravy, and poutine on your menu, and they come in all different sizes, a person can come up to the drive-thru window and say ‘can I have a large fry’, or ‘can I have a large fry with gravy’, or ‘can I have 3 fries, and one of them has poutine’ – you have to train the computer to understand this type of order.

Instead of using ASR (Automated Speech Recognition Software), the data must be formatted according to a very specific style guide provided by the company in order to achieve the highest accuracy needed to train the computer.

For example, assume you have a 45-second piece of audio with a person saying, “Could I get like, uh, a large fry and a medium fry with gravy, and a poutine?” When you type that, a transcriptionist writes that the Large Fry has a ‘capital L’ at the beginning and ‘Fry’ has a capital F at the beginning to indicate a specific order item. 

The company that wants this done would write a script for the computer to understand the various tags/keywords. Our team will be aware of the format to use because the company will inform us of the required formatting, and then our team will transcribe and adhere to the required style guide to format the data for the company’s software systems.

It depends on the task you’re attempting to teach your computer. Assume that all you want your AI to do is route the request to the appropriate person in a call center. The amount of data required will be determined by how many options a customer may have when calling a call center.

Let’s say you call your bank, and you get put into the IVR Phone System – ‘Please tell us the department you’re looking for, press 1 for accounts, press 2 for a credit card, Press 3 for payment.’ The problem is that customers calling into that IVR, say the same thing to get the same answer – for instance, if someone wants to open a credit card and press option 2, they may say ‘credit card’, they may say ‘i want to start a credit card’ or ‘I’d like to apply for a credit card’ or ‘credit card please’ and our team would format the data for the AI to understand those different phrases, which all mean the same thing, and know how to direct the call.

The best answer boils down to one fundamental question: how many different ways can the same thing be said for the computer to complete the task? Thousands of hours of data are typically required. It’s similar to teaching a child how to speak and think. How many times do you have to tell your dog to sit before he will? It takes practice in a variety of settings.

Thousands of hours of real-world data with context. Here are some examples of data types with lower quality.

Example #1 – Trimmed Data: Some of our clients want to trim their data so they don’t have to pay to have too much transcribed. They’ll listen to the audio of someone placing an order, so say someone orders for 1.5 minutes, but they only want to know where the person said ‘fries.’ They’ll cut out the beginning and end of the conversation, so our transcriptionists won’t understand the context of the conversation when we get the audio, so we need real data with context.

Example #2 – Synthetic Data: Some companies will provide us with fake data where they’ll record themselves ordering in as many different ways as they can think. It’s not true data, it’s not real data – and what they get back is that the computer will understand how to sort their orders but no one else since it’s not being done in the true context of the environment. In certain situations, it’s ok to use fake data – but in most cases, it’s not – for instance, if you’re trying to train AI about medical interviews and you want the computer to pick out ‘diagnosis’ or ‘symptoms’, you have to train the computer how to identify a diagnosis or a symptom, but you can’t use real data because it’s against the law and violates HIPAA.

Companies like Microsoft and Google offer and sell thousands of hours of generic conversation and datasets as well – and they use real data. There are pros and cons to this depending on the data available. You could find thousands of hours of ‘men speaking’ – so if you’re trying to train a machine to identify if a man is speaking, that could be a realistic use case – but not much more than that.

• Expertise combined with a human touch
• Flexibility in relation to the style guide and requirements
• Security includes multiple safeguards to protect the privacy of personal information
• Workforce size to quickly scale projects