Detecting Internal and External Noises to Improve Driver’s Safety

Services Provided: Audio Annotation

The client was developing a new AI tool that would be able to recognize all kinds of sounds inside and outside a car which would alert drivers about a potential danger. In order to train the AI tool, the client prepared training data in the form of video with audio of various sounds. All of the events heard in the audio needed to be identified and localized. This included things like the speech of different people, sounds inside the car, sounds outside the car, gadgets, pets, etc. All of this needed to be done with pinpoint accuracy

To train their AI system, the client was looking for a vendor who would be able to recognize all of the sounds heard in the videos and localize them with millisecond accuracy and identify violence, aggression, and other emotions. These tasks were not as straightforward as they may seem since there were many different sounds coming from various sources and were overlapping with one another. Some events needed to be recognized in the video and audio, but the latter would serve as the source of truth for the AI system. A high level of attention to detail was required to complete these tasks. Otherwise, a sound could be missed, which would lower the quality of the end product. 

Since the client is based in Germany, the audio in the videos was also in German. The customer had a strict deadline of 3 months to complete all of the annotation work, which meant that all of the sourcing and recruiting, training, and other project actualization processes needed to fit into that timeframe as well. 

The client announced a tender with a very strict set of requirements for the winning company. The vendor needed to have extensive experience with audio annotation and be able to achieve a quality score of 98%+. Since the process of manual audio annotation would be time-consuming, the client was asking all of the tender participants to present their ideas on how they could optimize the time that would need to be spent on annotation.

In the end, Mindy Support proved to be victorious based on the following criteria:

• High quality of work performed during the POC stage
• Estimated time frames to complete the project
• The thorough training process for all data annotators 
• The costs of the annotation work

As mentioned earlier, the client had a strict deadline of 3 months for when all of the work needed to be completed, so we knew we had to act fast. Thanks to our size and experience implementing projects of various sizes and complexities, we were able to hire a team of 45 annotators to work on the project. Once we assembled our team, we started the training process right away since the team had to have a consistent understanding of all sound categories as well as attributes such as aggression, anger, positive reactions, etc., which could range from high to low. The quality of the training was critical since the consistency of the annotations was dependent on it. 

Since our team members completed many audio annotation projects in the past, it did not take them long to gain an understanding of how the client wanted all of the audio annotations to be done and how to deliver the needed quality levels within a short time period. In fact, it was this experience that allowed us to identify an area of improvement which could significantly speed up the annotation process. The client requested that we use an open-source tool to complete the annotation work, but after conducting an analysis of the platform, we understood that it does not suit the tasks at hand and suggested using another tool to complete the work. This insider knowledge allowed us to increase the pace of annotations by 30%. 

In the end, we delivered a customized output to the client, so they could continue to use their platform for internal checks. We conducted multiple checks on our side to make sure the required quality level was met. In the end, we completed the project within the specified timeframe and quality levels. The client was so satisfied with our work that they continued to expand the project with updated guidelines 6 months after the first phase was completed. All in all, we have been working on this project for 2 years in total, with excellent feedback from the client. 

• Assembled a team of 45 data annotators
• Accelerated the annotation process by 30% thanks to expert knowledge and experience
• Completed the core part of the project within a strict timeframe of 3 months 
• Successful cooperation with additional project phases during 2 years