Issues

• Harden the scoring harness: switch-point metrics and a normalisation audit

  evaluation
  RahilOp/liquid-ai#11 · Filter by author Awshesh12 opened on Jul 22, 2026
  

• Real-time streaming inference on edge devices

  deployment
  RahilOp/liquid-ai#10 · Filter by author Awshesh12 opened on Jul 22, 2026

• Extend to trilingual code-switching (JA-EN-X)

  dataresearch
  RahilOp/liquid-ai#9 · Filter by author Awshesh12 opened on Jul 22, 2026

• Expand the augmentation pipeline beyond MUSAN noise

  data
  RahilOp/liquid-ai#8 · Filter by author Awshesh12 opened on Jul 22, 2026
  1

• Mix real code-switching corpora into training, not just evaluation

  datatraining
  RahilOp/liquid-ai#7 · Filter by author Awshesh12 opened on Jul 22, 2026
  1

• Build a speaker-diverse real-speech evaluation set

  evaluationresearch
  RahilOp/liquid-ai#6 · Filter by author Awshesh12 opened on Jul 22, 2026
  

• Version-control the real-speech predictions so Table 5 is reproducible

  evaluationpaper-blocker
  RahilOp/liquid-ai#5 · Filter by author Awshesh12 opened on Jul 22, 2026
  1

  

• Disentangle LoRA rank from data mixture in the final ablation step

  paper-blockertraining
  RahilOp/liquid-ai#4 · Filter by author Awshesh12 opened on Jul 22, 2026
  1

• Score the merged Whisper checkpoint on the FLEURS monolingual controls

  evaluationpaper-blocker
  RahilOp/liquid-ai#2 · Filter by author Awshesh12 opened on Jul 22, 2026
  1

  

• Recover the training recipe for the best LFM run (r32 encoder LoRA + FLEURS mix)

  paper-blockertraining
  RahilOp/liquid-ai#1 · Filter by author Awshesh12 opened on Jul 22, 2026