Case Study

Machine Transcription Training for 30+ Languages

One of our clients, a well-known language service provider (LSP), was struggling to meet transcription output targets for a large-scale machine transcription training project that needed to be transcribed in over 30 languages.

Since there are very few transcription service providers who can handle quality transcription output for that many languages, they had to resort to hiring freelancers to be able to deliver on the project. However, the associated cost of hiring and training freelancers was starting to become expensive for them.

With Audio Bee, the LSP was able to deliver continuous outputs for their client without the hassle of hiring and training freelancers, in a cost effective way.

The Challenges

Lack of Trained Resources

While transcription is a mature industry for English and popular Western European languages, there is a lack of trained resources for other languages due to rarity or low demand. Thus, most of the LSPs cannot easily increase production at an affordable cost.

Quality & Scale

The company needed to deliver fast output but they did not have the luxury of training each and every transcriber. Even when they hired native workers who could transcribe fairly accurately, the outputs were still not meeting the stands their client had demanded. Since there is a certain learning curve for workers to be able to start producing quality transcripts, simply hiring native speakers does not guarantee quality results.

Handling Annotation Requirements

Multilingual machine training transcription work is different from regular transcription jobs. It requires transcribers to follow client provided Word Domain Convention (WDC) guidelines, which are generally lengthy and complicated even for native speakers. For example, the transcribers are required to follow annotation and speaker segmentation rules accurately to train machine transcription technologies.

The Solution

Seamless integration with the client’s work processes

Tools agnostic approach for work harmony and faster outputs

Hybrid workforce consisting of specialist annotators and native transcribers

The Results

Audio Bee’s trained resources with prior experience in machine transcription guidelines were perfectly suited to tackle this project from the get go. Our understanding of annotation and speaker segmentation rules for such clients allowed us to deliver quality output efficiently.

Quality Output, Fast

The company was able to see quality output within the first week of starting the project despite our resources having to pass their testing procedure, which had delays. Initially, we took up to two weeks on the delivery of the first couple of tasks in some languages. However, by the time we were on the tenth language, we dropped it down and maintained a consistency of 1-3 days depending on the task length and complexity.

Quick Learning to Improve Errors

Our team was able to clearly perceive any feedback on errors, communicate them with our transcribers, and make sure not to repeat them in future tasks. Our annotation team is highly experienced in this since they operate across multiple languages and, thus, we were able to quickly incorporate quality improvement feedback. Every client has their own nuances and we understand them much better due to our experience working with machine transcription training technologies. The company interpreted some rules differently and we were able to easily incorporate them in our tasks.

Scalability of the Team

The company wanted us to continue increasing our output for those languages for which they had a lot of data. They became heavily dependent on our team for delivery. It was difficult to find good resources in some of these languages due to multiple reasons, yet we found new ways to attain them. We kept improving our training processes to get them producing good quality output fast.

Sign up to Audio Bee today!

img

Contact Us

sales@audiobee.ai
fb
twitter
linkedIn

SUBSCRIBE TO OUR NEWSLETTER

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

Copyright © 2021 Audio Bee, Inc. All rights reserved.