R-TASA × SpeechTransformer (PyTorch)

PyTorch implementation of Residual Transmitted & Aggregated Self-Attention (R-TASA) integrated into SpeechTransformer for ASR.

📄 Zhang, Han, et al. “TASA: Transmitted and Aggregated Self-Attention for Speech Recognition.” INTERSPEECH 2024. [paper]
📄 Dong, Linhao, et al. “Speech-Transformer: A No-Recurrence Sequence-to-Sequence Model for Speech Recognition.” ICASSP 2018. [paper]

---

🛠️ Install

git clone https://github.com/itsmekhoathekid/TASA.git
cd TASA
pip install -r requirements.txt

---

⚙️ Usage

Enable R-TASA in config:

model:
  attention_type: "r_tasa" 

Train example:

python train.py --config config/r_tasa_local.yaml

---