podcast-dramabox-dacvae-pairs

Paired audio-codec latents for training a DramaBox → DACVAE latent translator. Both codecs share an identical grid: 25 Hz, 128-dim, frame-aligned (same length).

Derived from TTS-AGI/podcast-tokenized-bg3.5-enj5.

How it was built (per sample)

DACVAE latent (from source dataset, = target) → DACVAE.decode → 48 kHz mono wav → duplicate to stereo → DramaBox/LTX-2.3 audio VAE encode → patchify → DramaBox latent (= input). Both latents trimmed to the min length (they differ by ≤1 frame).

• DACVAE: facebook/dacvae-watermarked (encoder_rates [2,8,10,12], latent_dim 1024, codebook_dim 128, 48 kHz).
• DramaBox: LTX-2.3 audio VAE from ResembleAI/Dramabox (dramabox-audio-components.safetensors), 16 kHz, 25 Hz latents.

Format

Each shard XX_XX_XXXX.npz (numpy.savez, load with allow_pickle=True):

key	type	description
db	object array of (T,128) float16	DramaBox/LTX latent — translator input
dac	object array of (T,128) float16	DACVAE latent — translator target
keys	(N,) str	sample keys (match the source dataset)
meta	(N,) str	JSON: transcript + emotion scores

import numpy as np
z = np.load("00_00_0000.npz", allow_pickle=True)
db, dac = z["db"], z["dac"]          # paired latents
x, y = db[0], dac[0]                 # (T,128), (T,128)