hyeyeon

DAC 16 kHz models (descript-audio-codec, commit c7cfc5d, source unmodified) retrained from scratch on speaker subsets of LibriSpeech train-clean-360.

Contents

p1/<model>/<checkpoint>/{dac,discriminator}/weights.pth

• models: ref_00 ... ref_05, target_00, target_01 (each trained on a different subset of 57 speakers, about 13 hours of audio)
• checkpoints: 50k, 100k, 200k, latest (= 250k steps)

p1/splits/ (which audio each model was trained on)

• models.csv: model names and roles
• speakers.csv: speakers with sex, minutes, chapters, role (pool or population) and held-out chapter
• membership_speaker.csv: 1 if a speaker's audio was in that model's training set
• membership_utt.csv: the same per utterance
• train_<model>.csv: the exact training file list of each model
• eval_set.csv: the shared evaluation utterances (group: candidate, holdout, population)
• manifest.json: split parameters and seed
• Paths are from the original machine; replace the LibriSpeech root with your own.

p1/configs/<model>.yml: the training config of each model (argbind format, for scripts/train.py)

Training

• 250k steps, batch 12, train crop 0.3795 s, otherwise the official 16 kHz recipe
• One NVIDIA B200, four models at a time, about 26.5 hours each

Quality (dev-clean, 500 files, 12 codebooks, latest)

STOI 0.968-0.970, PESQ-WB 3.64-3.73, mel L1 0.430-0.447. Official 16 kHz model on the same files: STOI 0.973, PESQ-WB 4.00, mel L1 0.364.

Loading

import dac
model = dac.DAC.load("p1/ref_00/latest/dac/weights.pth")