Streamlining your virtual assistant's speech with the highest audio quality
AI-powered virtual assistants require copious amounts of high-quality data to perform optimally.
The transcribed audio data needed is often produced through a process that typically falls outside a company’s core competency. That’s where TELUS Digital’s audio data annotation and labeling services can help.
The challenge
Our client is one of the world’s largest information technology companies, manufacturing consumer electronics products and providing online services.
Their devices feature an intelligent assistant that allows users to request and receive information through a dynamic speech interface in multiple languages. Since the client was looking to optimize their interface, a high-quality dataset that could train their AI-powered system quickly and effectively was required. Several challenges were inherent to this project:
- Maintaining transcription accuracy above 95%, no matter the scale of the project
- Staying within budgeted costs
- Securing sensitive data
- Providing ad hoc project support
- Meeting short turnaround time requirements
- Adhering to corporate branding standards
- Implementing data to enable live communication through machine learning
- Providing end-to-end debugging and testing
The TELUS Digital solution
Our client needed a single point of contact to kick-start the project and scale rapidly. It was imperative to generate fast results in a usable format, while maintaining a 95%+ standard in transcription accuracy.
TELUS Digital assembled teams of 3-15 native speakers for each of the 16 languages in scope to work in the client’s European and U.S.-based offices. To help with the physical requirements, we implemented high-level security and access restrictions to augment the client’s limited space and personnel. Project managers were also selected to ensure the transcribers perfectly understood the requirements of the project in order to deliver accurate results.
To create the needed dataset, our multilingual solution included:
- Transcription of spoken utterances
- Editing pre-transcribed text
- Annotation of the transcribed data for better quality audio and semantic content
- Services in several languages including, Arabic, Bulgarian, Chinese, Finnish, French, English, German, Greek, Indonesian, Italian, Nepalese, Romanian, Russian, Swedish, Ukrainian and Vietnamese
The results
TELUS Digital collectively compiled a team of 70 transcribers, who completed over 10,000 audio hours in just three months.
Throughout this process, the transcription team maintained over 95% accuracy. As a result, the client launched their product release well ahead of their target deadline.
I am impressed by the team’s responsiveness, efficiency and subject
matter expertise that helped us solve problems throughout this complex value chain.
