Capacity-Matched Dense Marathi Conformer-512 (130M Parameters) 🎙️
A fully dense, capacity-matched Conformer foundation model for Marathi ASR, trained across 300,000 alignment steps to serve as the strict parameter-matched control (130.8M total parameters) against the dialect-conditioned Sparse MoE Conformer.
Architecture Highlights
- Base Architecture: 12-layer Dense Conformer ($d_{ ext{model}}=512$, 8 attention heads, FFN expansion factor = 8, kernel size 31).
- Parameters: 130,812,425 parameters (100% active per forward pass).
- Decoder: CTC greedy and beam decoding over 105 character vocabulary.
Benchmark Performance (RESPIN Held-out Suite)
- Malvani (D1): 45.29%
- Ahirani (D2): 42.41%
- Standard (D3): 31.33%
- Varhadi (D4): 39.90%
- Aggregate WER: 39.82% (CER: 11.40%)
- Downloads last month
- 45