Cochl. delivers top-quality machine listening technology to solve issues and challenges around the world. Based on research, the startup aims to create machine listening abilites that mimic human auditory perception. Datumo took part in the process of builiding an environmental sound recognition AI system based on deep learning, by collecting audio data from different places.
In order to build an AI model which recognizes a broader range of environmental sounds from residential areas, restaurants, subway stations, and more, there were some concerns involved. Let’s see how Datumo took care of this problem.
In order to prevent submission of artificial recordings, we implemented on-site recording feature to our crowd-sourcing platform. By carrying out the project through our mobile platform, crowd-workers were able to record the target sound anywhere, anytime.
Datumo and Cochl. tried to prevent edge cases by clarifying some standards beforehand. Some of the concerns were as follows:
1. Malls vs. Clothing stores with music playing
- Would it be possible to differentiate the music/noise coming from a mall and a clothing store? How would you characterize them?
- Unclear standards for validation
- Can you train AI models to differentiate the music/noise?
- Why is our client wanting to collect the noises separately?
2. Parks vs. Residential areas vs. Sreets
- Is there a particular sound/noise expected or not expected to be recorded?
- Should we have subdivisions within the “Parks” category? Would the sound/noise coming from small parks in the residential areas and large, national parks differ?
- Both residential areas and the streets will have sounds/noises coming from cars. What are some sounds/noises that we should try to avoid, depending on the location?
3. Homes
- For studio apartments, sounds may interfere with each other since there are no separations within the area.
- Would the inspector be able to tell the differences between the sounds coming from other areas(public toilets, offices, etc.) and the sounds coming from home?
- Specific standards must be set in order to make sure correct, targeted sounds have been collected.
4. Car horns
- Are all car horn sounds coming from the roads eligible?
- Does the length of the horn sound matter?
5. Playing music
- Can the music play louder than other noises?
- Do mobile ringtones and background music playing on the television all count as “Playing music”?
We thought there was a possibility of having private information included in the recordings, whether intentional or not. Therefore, we stated warnings in the project tutorials that recordings that include any form of privacy invading information would not be accepted. Also, during the final validation process, we carried out an extra step of filtering and deleting such data.
Cochl. is:
- 2-year consecutive winner of IEEE DCASE ’17 ’18, the most prestigious competition in sound AI (ranked first among 558 teams)
- The only Korean company to have been nominated as one of the Top 15 of Slush 100
- One of the top 4 AI startups in autonomous systems ranked by NVIDIA headquarters
- Giving speeches at NVIDIA GTC Silicon Valley / SXSW
We are excited to see how Cochl. brings innovation to the industry.
