How Sage BPM helps a client bring down the word error rate to JUST 0.04 from 0.16?

THE BENIFITS

  • Sage BPM delivered 2 million data records, completely recorded, validated and tagged within 6 months of time.
  • The model accuracy improved at an impressive rate.
  • Within couple of months, the Word Error rate went down from 0.16 to JUST 0.04.
  • The Character error rate was at 0.03 which was high at 0.13

OVERVIEW

The client is the largest next-generation PoS platform for the F&B sector with more than 15,000+ clients across India. The company is present in more than 135 cities across India and UAE, working on a mission to organize the unorganized food sector in India and globally with the help of technology.

Business challenge

The client was working to build a speech recognition model for order taking and launch the SaaS model across all states of India. The challenge was to record millions of data (consisting of dishes, phrases used in conversation while order taking) in local accent in male as well as female voice for all the states of India, and deliver validated as well as tagged data ready for the model to be trained. The candidates would need to record the data in their natural local accent, not neutral.

How Sage helped?

Sage BPM analyzed the internal team first to figure out who can contribute for which state. For the rest, Sage BPM reached out for freelancers, belonging to different regions/states of India. Once the recording was done, a QA from the Sage BPM team would check the data for accuracy, and give feedback on a daily basis, get the data re-worked where required. Once the data is recorded accurately, there would be validation in client-owned software, which ensures data is ready for the model to be trained. Simultaneously, the Sage BPM team would classify/tag each data record/dishes in a particular bucket as per the respective region/state such as main course, starters, snacks, break-fast, dessert, and so on.