Local Accent Speech to Text Transcription Engine
Implementation Time:
9 months
Solution Provider: AI Singapore
The Ministry of Social and Family Development is a ministry of the Governement of Singapore focusing on nurturing resilient individuals, strong families and a caring society in Singapore.Â
- MSF is currently adopting Natural Language Processing (NLP) technology to answer simple queries in the Baby Bonus Hotline called Ask Jamie Voice (AJV)
- Analysis shows that AJV’s speech to text transcription is only about 56% accurate, and 35% of the transcriptions in a sampling of 280 call records contained errors
How can MSF develop a speech to text transcription engine that can better recognise local accent, account for Singlish and Code Switching, and is customisable to MSF’s terminology?
The AI Speech Lab (launched by AISG and led by Prof Li Haizhou (NUS) and Prof Chng Eng Siong (NTU)) has an Automatic Speech Transcribing system that could interpret and process the unique vocabulary used by Singaporeans – including Singlish and dialects – in daily conversations.
- A multitask learning (MTL) framework using language identification (LID) as the auxiliary task was employed to help improve the code-switch speech recognition perform
- A word vocabulary expansion method was applied to alleviate cross-lingual data sparsity issue in language modelling
- The Asterisk framework with Nautilus was employed to demonstrate the ability of the system to intercept real phone calls and perform real-time transcription
Outcomes
- Developed algorithms to train and adapt acoustic and language models for Large Vocabulary Continuous Speech Recognition (over 60 thousands words) to transcribe into Singlish and code switch between English and Mandarin
- Developed a system that can transcribe live speech automatically with sentence breaks and speaker turns, approximating human listening performance
- Word error rate (WER) decreased from 33.47% to 15.67% for one test set; and from 37.30% to 14.96% for another test set. The overall accuracy of the engine is 84.69%.Â
- Worked with the existing chatbot provider to integrate the engine into the chatbotm allowing it to run in parallel with the current Google Speech API solution
- Automatic Speech Transcribing system could be deployed at various government agencies and companies to assist frontline officers in call centre-type of work
PDF Document
This is the last content of this tab. If you do not see any resources above, it means the solution provider have not provided any resources. Feel free to contact Solution Provider for more information.
Implementation Time
9 months
Use Case Brochure
