← Back to portfolio

Hyundai Motor Company AIR Lab · Feb 2022 – Apr 2022

In-vehicle & Mobile Speech Recognition

Role · Speech-recognition pipeline research, LM fine-tuning

01Background & Goals

  • Engine and road noise degrade speech recognition inside a car.
  • To offset this, we used a language model in decoding, then proposed and studied decoding with a language model that reads the previous dialogue turns.

02Contributions

Context-aware LM-integrated decoding
  1. Trained a generative LM (KoGPT) on spoken-dialogue data
  2. Modified the decoding algorithm to combine acoustic-model and LM scores
  3. Built a recognition pipeline that includes the previous dialogue turns

03Tech Stack

Framework / Platform
PyTorch, Fairseq, Transformers
Methodology
Joint decoding, Beam search, LM fine-tuning

04Results

  • Improved recognition accuracy by 5% under external noise