Mercury-M18nc

A full-duplex speech-to-speech model fine-tuned for time-aware conversation. Released under a codename while the associated paper is under anonymous review.

What this checkpoint is

PAD loss weight 1.0, effective batch 1, lr 1e-6, no native clock. Best checkpoint in this series.

Trained for 2 epochs on 4,440 self-generated full-duplex dialogues whose word timestamps come from forced alignment of the rendered audio (not from the model's own decision frames).

Reported measurements

Text/audio consistency on 100 held-out dialogues — does the produced audio actually say what the text-decision stream decided? Both sides come from the model's own artifacts; no reference transcript is involved.

	this checkpoint	untuned base
character-level consistency (mean / median)	88.9% / 91.2%	92.6% / 96.2%
word-level recall	72.7%	79.1%
severe mismatch (<50%)	4/100	6/100
text emission rate	3.2 tok/s	2.5 tok/s

Frames are 80 ms (12.5 fps), audio is 24 kHz.

Attribution and licence

Derived from a publicly available full-duplex speech model released under CC-BY-NC-4.0. This checkpoint inherits that licence: non-commercial use only. The base model's name and full attribution are withheld only for the duration of anonymous review and will be added here once the associated paper is de-anonymised.

Not included

Training code, data, and evaluation scripts are withheld during anonymous review.