Datasets:

TTS-AGI
/

additional-data-a1-plus

MOSS-Audio-Tokenizer-v2 codes (12-codebook RVQ @ 48 kHz) with captions for scientifi-papers/a1-plus, ready for MOSS-TTS-Local-Transformer (moss 1.5 local) training. No MP3s — tokens + text + captions only.

tokenized_audio.parquet, per row (key joins to a1-plus):

target_codes : int16 bytes, reshape [target_frames, 12]
ref_codes : int16 bytes, reshape [ref_frames, 12] (same-speaker reference; may be null)
text : transcript (what is said)
voice_acting_caption : natural GENERAL + SCRIPT caption with [pause X.Xs] (how it is said)
procedural_caption : terse tag-style caption
bude_caption : BUD-E voice description

import numpy as np, pandas as pd, torch
d = pd.read_parquet("tokenized_audio.parquet")
r = d.iloc[0]
tgt = np.frombuffer(r.target_codes, np.int16).reshape(r.target_frames, 12)   # [T,12]
prompt = r.voice_acting_caption          # instruction/caption for training
# decode: MossAudioTokenizer.batch_decode([torch.tensor(tgt.T)])  -> 48kHz audio

Full VoiceNet/EmoNet/blend/genuineness scores + same-speaker ref_spk_sim live in the scientifi-papers/a1-plus metadata (password acting). Reconstruction verified at 0.994 correlation.