Santiago Valdarrama
@svpino
Streaming gains extend to 18 test sets across 10 additional languages, all improving over the base model.
Four areas I'd like to see more tests:
1. Self-consistent LSP: derive targets from the streaming model's own outputs to potentially move beyond the base model's ceiling.
2. Adaptive chunks: vary chunk size with audio complexity to improve the latency–accuracy trade-off.
3. Continuous rewards: add confidence estimates to improve calibration and streaming decisions.
4. Encoder adapters: test lightweight LoRA adapters for specialized domains.
