user
Explain the following
---
assistant
ea ("Feed"): Modal seq2seq, mel-in / text-out. Vocab is BPE-32k, all langs. Enc: 24× Conformer, d=1024, ker 15. Dec: 12× xformer, cross-attn to enc. RoPE. Trained w/ CTC aux loss (w=0.3) on enc top.
Positional: RoPE base 10k std. NB: for >30s audio we extend base to 500k (NTK-aware, see Peng et al.) — do NOT retrain, just swap at inference. Works b/c enc is conv-heavy.
Bill: batches of 64 utt, dyn padding to longest in batch. Bucketing by len (8 buckets). LR 1e-3, warmup 10k, cos decay. AdamW b2=0.98. Grad clip 1.0.
Sunny: WER on ls-clean 2.1, ls-other 4.7. CV-en 8.9. Aishell-1 4.2 (CER). Not great on code-switch, ~15% WER. Fix w/ more CS data next qtr.
---
Anything so I understand it.