Yesterday
-
Data step builds from the public datasets
-
Scripts under the duplex package name; card final pass
-
Card: text accuracy against speech
-
Card: drop the note on removed folders
-
Remove checkpoints from earlier runs on older data
-
Real-speech ceiling for the two-track layout
-
Benchmarks and clips with the true prompt text
-
Add training scripts and Apache 2.0 licence
-
Rename to duplex-0 and report the retraining only
2 days ago
-
Training report for the 17 November retraining of the three options
-
resplit-20261117/conversational-two-best: step 151
-
resplit-20261117/conversational-two: step 301
-
resplit-20261117/pretraining-two-best: step 14220
-
resplit-20261117/pretraining-two: step 17064
-
resplit-20261117/conversational-one-best: step 139
-
resplit-20261117/conversational-one: step 278
-
resplit-20261117/pretraining-one-best: step 17849
-
resplit-20261117/pretraining-one: step 20595
-
resplit-20261117/conversational-none-best: step 229
-
resplit-20261117/conversational-none: step 229
-
resplit-20261117/pretraining-none-best: step 11160
-
resplit-20261117/pretraining-none: step 13392
-
Remove listening prompt 1 (music bed and distorted speaker tracks)
-
Add the 17 November 2026 training report, plots and listening samples
-
stop-20261117/conversational-two-best: step 153
-
stop-20261117/conversational-two: step 305
-
stop-20261117/mid-training-two-best: step 570
-
stop-20261117/mid-training-two: step 570
-
stop-20261117/conversational-one-best: step 141
-
stop-20261117/conversational-one: step 282
-
stop-20261117/mid-training-one-best: step 529
-
stop-20261117/mid-training-one: step 529
-
stop-20261117/pretraining-two-best: step 15653
-
stop-20261117/pretraining-one-best: step 17862
-
stop-20261117/pretraining-two: step 18499
-
stop-20261117/pretraining-one: step 20610
-
stop-20261117/conversational-none-best: step 116
-
stop-20261117/conversational-none: step 232
-
stop-20261117/mid-training-none-best: step 217
-
stop-20261117/mid-training-none: step 434
-
stop-20261117/pretraining-none-best: step 13392
-
stop-20261117/pretraining-none: step 15624
-
full-20261117/mid-training-two-control-best: step 11046
-
full-20261117/mid-training-two-control: step 11046
-
full-20261117/mid-training-one-control-best: step 10178
-
full-20261117/mid-training-one-control: step 10178
-
full-20261117/mid-training-two-best: step 553
-
full-20261117/mid-training-two: step 11046
-
full-20261117/mid-training-one-best: step 509
-
full-20261117/mid-training-one: step 10178