Voice Assistant SKU Training Simplified - The What, Why, and How of it all.

If you are planning on building a Voice Assistant using off-the-shelf services available online, you might be in for a rude awakening. These services are simply not equipped to handle the complexities that need to be dealt with to give customers a truly seamless Voice experience.

A Voice Assistant before it can truly be accurate and effective needs to be trained for your domain(eg Retail or Travel) and subdomain(Grocery or Flight). 

Training however is resource-intensive and not just a one-time effort. Voice Assistants need to be trained continuously to recognize the various SKUs that your online store offers and trained again as your store grows.

But wait, what exactly does training mean? And why is it so important?

Well, let’s break things down so it’s easy to understand -

Voice Assistant SKU Training

1) ASR Domain Optimizations

Imagine this, Tim is launching a brand new Voice Search feature in his car reselling app for his consumers. He believes Voice is growing exponentially, and adding a Voice Assistant to his app will undoubtedly be very beneficial.

Tim is elated once the feature has gone live, but soon the analytics start pouring in.

‍

‍

Tim’s excitement slowly begins to fizzle out as more such issues begin to arise. What the hell is happening?” he exclaims.

Speech to text services convert speech captured via a microphone on devices to text in the language spoken by the user. However, what happens is that Voice Assistants do not understand in what context the user has spoken terms, which can lead to errors.

The customer searched for “Toyota Corolla” in Tim’s case, but since Tim used an off-the-shelf generic ASR, the Assistant recognized the term as “Toyota Gorilla”. The Assistant simply does not understand that for Tim’s app domain “Corolla” is a term that is more relevant and “Gorilla” does not have a place here.

‍

We do pre-training to ensure that errors such as this do not occur. In cases where several similar-sounding words are recognized as potential searches, the Assistant must know which term to pick so that customers are shown results for precisely the term they intended to search for.

When we train our Voice Assistants, we ensure that SKUs that belong in the domain are assigned a higher probability (or weight); and, because of this, are recognized over terms that sound similar but are irrelevant in that context. Customers now can search for “Corollas” and not be shown “Gorillas”

‍

2) Translation Optimizations

Sally runs an online grocery store app where she sells fresh produce from her garden direct to consumers online. She too wanted to add a Voice Assistant to her app to make ordering fruits and veggies a breeze for her customers. 

Sally was having great success with her Voice Assistant but upon looking at the analytics she noticed that there were several queries for terms such as brinjals, aloo, chikoos, and coriander that weren’t showing any results to consumers.

‍

“Why is this happening? I’m losing so much business” she mutters.

Well Sally, although your SKU name might be “Coriandor 200gms” the Voice Assistant in the app must understand that the item must be pulled up even if a customer searches for “Parsley” since they are referring to the same item. 

This is the case for many SKUs, several of them have synonyms and are called by different names. 

 For example in India, an Eggplant or Aubergine is more commonly referred to as a Brinjal. Therefore,  Sally’s Assistant must be able to recognize this synonym and understand that if a customer is looking to order 2 kgs of “Brinjal” she is, in fact, looking to buy some “Eggplants” or “Aubergine” and vice versa.

‍

‍

And in many cases even if a query is made in English, the SKU name might be spoken in the local language. A query such as “Add 2kgs Aloo to my cart” is not uncommon.

‍

A generic translation engine lacks the context in which a word in the sentence has been used and can give unreliable translations for queries. These generic translation services need to be optimized for every domain. For different synonyms and as well as local languages.

A Voice Assistant must be trained well so it may handle these intricacies.

Pre-trained SKUs

To make your life easier

So far it seems that training is a long and arduous process, and building a Voice Assistant seems highly impossible and perhaps simply not the effort. Well, just hold on for a second.

At Slang Labs we understand that building and then having to train a Voice Assistant can be a difficult task, which is why Slang In-App Voice Assistants come pre-trained with thousands of SKUs for various domains.

Making the decision to integrate a custom Voice Assistant to your mobile or web apps that much simpler.

The Retail Domain

Slang In-App Voice Assistant comes pre-trained for over 45,000+ SKUs in the retail domain. 

This means that subdomains that fall under the retail domain, such as grocery or pharmacy, upon integrating Slang Multilingual Voice Assistant, can provide customers a Voice Assistant to use straight out of the box.

Slang In-App Voice Assistants are pre-trained for 35,000+ Grocery SKUs, here’s a table displaying a few of these -

‍

‍

And pre-trained for 10,000+ Pharmacy SKUs such as -

‍

‍

We aim to cover most common SKU’s that fall under these domains, ensuring that once you integrate Slang Retail Voice Assistant you can begin using it almost immediately.

‍

The Travel Domain

Slang In-App Voice Assistant also comes pre-trained for thousands of city and town names across India.

Proper nouns can be tricky for Voice Assistants to accurately recognize, the assistant must take into account different pronunciations and translations. We do the hard work of training the assistant to recognize these names efficiently and accurately. 

What this ensures is that the amount of work you need to do post-integration is greatly reduced, saving you both time and money.

Slang In-App Voice Assistant is pre-trained to recognize close to 130 cities for the flight subdomain -

‍

‍

And close to 45,000 city and town names in the bus subdomain such as -

‍

Integrating a Voice Assistant for travel apps has never been easier.

‍

More about our Knowledge Graph

We understand that as time passes and businesses grow so do the number of products that need to be listed on retail apps. The same is the case for cities and towns in travel apps.

We ensure that our list of SKUs for various domains is continually growing and evolving. Every month new data is added to our existing datasets to keep them up to date. We don’t use any data provided by our customers to augment our global knowledge graph. We work with trusted third-party partners to discover and add new SKUs to our knowledge graph continuously.

Furthermore, we also keep a close eye on search patterns and if we detect certain SKUs not being found, we fix the error, thereby increasing accuracy over time. All this is managed by Slang and requires no intervention from our customers.  

But the question arises, what if my SKUs are not in your global knowledge graph? Don’t worry we have you covered. ‍

Custom SKU’s

Yes, pre-trained SKUs are great, but what if your business requires custom SKUs to be added?

With Slang you can choose to use our existing pre-trained SKU list, upload a list of your own custom SKUs, or even employ a combination of both.

Does this sound complex? Well, you don’t have to worry. Adding custom data to the Retail and Travel subdomain with Slang CONVA, the world’s first In-App Voice Assistants as a Service platform is a breeze. 

And the Custom Knowledge Graph that holds your custom SKUs and is limited to your business. It resides within your account and is not shared with our Global Knowledge Base. 

‍

‍

Once you begin to use our console for each subdomain in the retail and travel domains, you can upload the following data files:

• SKU/ Location data
• Filters data
• Synonyms data
• Language data

Additionally, in ‘custom/miscellaneous’, we can add the following data -

• Navigation data

By adding these simple parameters, we can ensure that the Voice Assistant you meet is tuned to be more intelligent, accurate, and efficient.

It’s just that simple.

To learn more about how exactly you can customize the retail and travel subdomain data click here.

Conclusion

With pre-trained SKUs and the ability to add custom SKUs, it has never been easier to bring the Voice Assistant experience to your customers! We go above and beyond to ensure that once you integrate our Voice Assistant it is functional and usable right out of the box.

At Slang Labs we’re committed to our vision of a Voice Assistant in every app, to learn more about us and other steps we’ve taken to make this a reality click here.

To have a more comprehensive conversation book a demo with us today!‍