Multilingual Data creation for Japanese Language – Case Study

 About Client:

Client is a AI Technology Company that builds fully autonomous SaaS products for our global customer base using the latest AI software technologies. They are building game-changing Artificial Intelligence Technology Solutions that are easy to use, adapt, and scale, making our clients successful in a fully connected world.

Requirement:

Recordings in –

1. Six different accents of the Japanese language (Kanto, Tōkai–Tōsan, Kinki/Hokuriku, Hokkaido/Tohoku, Chugoku/Shikoku and Kyushu), according to Kanto standards as it will help the speech recognition model .
2. Each unique speaker between the ages of 15 to 60 years old with gender equality amongst the speakers.
3. The recordings should be around 100 speech hours per accent, with each session not exceeding few hours.
4. Each conversation should consist of two speakers and each conversation should end between 5 to 15 minutes.
5. The recording method should be stereo audio with two channels in .wav file format, with one single channel per speaker.

Challenges:

1. Speakers requirement:
   The number of resources required for the entire project were around 500 speakers with an equal male-female ratio & age ranging between 18 to 60 years old.
2. Recording method:
   The requirement was to record  speakers per conversation, one speaker per channel in a quiet recording environment.
3. Audio Editing:
   Removing silence and poor audio parts in the recordings.
4. Finding right speakers:
   Finding so many speakers of different language and arranging their schedule for session.

Solution:

• Fidel was able to find and onboard more than 500 speakers of defined accents within a very short period. A proper system was put in place where the speakers could register their details & timeframes and the system will then pair up the speakers for conversation recordings based on the data.
• The entire project, including recordings, Quality checks, detaied listing of Meta data and transfer of deliverables were completed within the short schedule of 2 months.
• We provided 20% more recording, in total hours of recordings.

Result:

• This dataset helped our client in building speech recognition model .
• Client’s business revenue was increased because the project helped them enter a new market and establish their brand.
• This helped them create a broader user base and enhance business volume.