Multi-Teacher Knowledge Distillation for Accented English Speech Recognition

• Master's Thesis:
• Presentation Slides: ./MTKD for AccentedASR.pdf
• Oral Presentation: https://www.youtube.com/watch?v=n5lFmgq_Fy0
• Publication I @Interspeech 2025: https://arxiv.org/abs/2506.08717
• Publication II:

🖼️ System Diagram of the Proposed MTKD Method

✨ Features

• Multi-Teacher Distillation for robust ASR
• Support for various teacher weighting strategies:
  • Average (mtkd_avg)
  • Top-1 (mtkd_top1)
  • Rank-Weighted (mtkd_rankweighted)
• Accent specific fine-tuning baselines
• Logits extraction from teacher models
• Easy-to-configure training and evaluation pipeline

⚙️ Installation

1. Clone the repository

   git clone https://github.com/aalto-speech/MTKD4AccentedASR.git
   cd MTKD4AccentedASR

2. Create and activate the virtual environment

   conda env create -f environment.yml
   conda activate accented_asr

🚀 Usage

🪄 Extract logits

python main.py \
  --TRAIN_ACCENT "American" \
  --DEVEL_ACCENT "Indian" \
  --TEST_ACCENT "Canadian" \
  --BATCH_SIZE 16 \
  --LEARNING_RATE 0.0001 \
  --N_EPOCHS 20 \
  --TRAINING 0

🧪 Train and Evaluate the Fine-Tuned Baseline

python ./ft/main.py \
  --TRAIN_ACCENT "American" \
  --DEVEL_ACCENT "Indian" \
  --TEST_ACCENT "Canadian" \
  --MODEL_CKP "facebook/wav2vec2-base-960h" \
  --BATCH_SIZE 8 \
  --LEARNING_RATE 0.0001 \
  --N_EPOCHS 20 \
  --TRAINING 1

python ./ft/main.py \
  --TRAIN_ACCENT "American" \
  --DEVEL_ACCENT "Indian" \
  --TEST_ACCENT "Canadian" \
  --MODEL_CKP "facebook/wav2vec2-base-960h" \
  --TRAINING 0

🧪 Train and Evaluate the proposed MTKD Method

python ./mtkd_avg/main.py \
  --TRAIN_ACCENT "American" \
  --DEVEL_ACCENT "Indian" \
  --TEST_ACCENT "Canadian" \
  --BATCH_SIZE 8 \
  --LEARNING_RATE 0.0001 \
  --N_EPOCHS 20 \
  --TRAINING 1

python ./mtkd_avg/main.py \
  --TRAIN_ACCENT "American" \
  --DEVEL_ACCENT "Indian" \
  --TEST_ACCENT "Canadian" \
  --TRAINING 0

🛠️ Customization

• --TRAIN_ACCENT: American, Canadian, Indian, Korean, Russian, British, Chinese, Japanese, Portuguese, Spanish, AESRC200H, EDACCdevel, AESRC200H+EDACCdevel
• --DEVEL_ACCENT: American, Canadian, Indian, Korean, Russian, British, Chinese, Japanese, Portuguese, Spanish
• --TEST_ACCENT: American, Canadian, Indian, Korean, Russian, British, Chinese, Japanese, Portuguese, Spanish, AESRC20H, AESRC10Accents, AESRC10Accents+AESRC20H, EDACCtest
• --MODEL_CKP:
  • facebook/wav2vec2-base-960h
  • facebook/wav2vec2-xls-r-300m
  • facebook/wav2vec2-base-100k-voxpopuli
  • elgeish/wav2vec2-base-timit-asr
  • patrickvonplaten/wav2vec2-base-timit-demo-colab
  • AKulk/wav2vec2-base-timit-epochs20
• --N_EPOCHS: Integers > 0: z ∈ ℤ, z > 0
• --BATCH_SIZE: Powers of Two: z = 2ᵏ where k ∈ ℤ, k ≥ 0
• --LEARNING_RATE: Positive Reals: x ∈ ℝ, x > 0.0

📚 Citation

Publication I

@article{bijoy2025multi,
  title={Multi-Teacher Language-Aware Knowledge Distillation for Multilingual Speech Emotion Recognition},
  author={Bijoy, Mehedi Hasan and Porjazovski, Dejan and Gr{\'o}sz, Tam{\'a}s and Kurimo, Mikko},
  journal={arXiv preprint arXiv:2506.08717},
  year={2025}
}

Publication II

...