TTS-AGI/commonvoice22-sidon-dacvae

CommonVoice 22 (Sidon-enhanced) converted to DAC VAE latents

Source

sarulab-speech/commonvoice22_sidon

Format

Each tar shard (~2GB) contains samples with three files per sample:

{sample_key}.audio.flac       # Original audio (FLAC, original sample rate)
{sample_key}.dacvae.npy       # DAC VAE latent [T_latent, 128] numpy float32
{sample_key}.metadata.json    # All metadata + duration_seconds + chars_per_second

DAC VAE Latent Format

• Model: mrfakename/dacvae-watermarked (Facebook DACVAE)
• Input sample rate: 48,000 Hz (audio resampled before encoding)
• Latent shape: [T_latent, 128] where T_latent = ceil(audio_samples / 1920)
• Latent rate: 25 frames/second
• Storage: numpy float32

Shard Naming

{LANG}-{split}-{index:05d}.tar (e.g., EN-train-00000.tar, DE-train-00001.tar)

Loading

With WebDataset

import webdataset as wds
import numpy as np
import json
import soundfile as sf
import io

url = "https://huggingface.co/datasets/TTS-AGI/commonvoice22-sidon-dacvae/resolve/main/EN-train-00000.tar"
dataset = wds.WebDataset(url).decode()

for sample in dataset:
    audio_bytes = sample["audio.flac"]
    latent = np.load(io.BytesIO(sample["dacvae.npy"]))  # [T, 128]
    meta = json.loads(sample["metadata.json"])
    print(f"Text: {meta['text']}, Duration: {meta['duration_seconds']}s, CPS: {meta['chars_per_second']}")

Decoding Latents Back to Audio

from dacvae import DACVAE
from huggingface_hub import hf_hub_download
import torch, numpy as np

model = DACVAE.load(hf_hub_download("mrfakename/dacvae-watermarked", "weights.pth")).cuda().eval()
latent = np.load("sample.dacvae.npy")  # [T_latent, 128]
z = torch.from_numpy(latent.T).unsqueeze(0).cuda()  # [1, 128, T_latent]
audio_48k = model.decode(z).squeeze(0).cpu()  # [1, T_audio] at 48kHz

Current Status

Shards uploaded: 935

Progress by Language

Language	Samples
AB_train	21,037
AF_train	139
AM_train	523
AR_train	28,531
AS_train	1,386
AZ_train	157
BA_train	121,197
BE_train	347,672
BG_train	4,952
BN_train	21,514
BR_train	3,510
CA_train	1,158,926
CK_train	7,878
CN_train	818
CS_train	21,731
CV_train	1,456
CY_train	8,014
DA_train	5,699
DE_train	607,871
DY_train	88
EL_train	1,934
EN_train	1,138,759
EO_train	128,103
ES_train	353,699
ET_train	3,402
EU_train	130,043
FA_train	29,789
FI_train	2,093
FR_train	593,066
FY_train	3,924
GA_train	546
GL_train	70,039
GN_train	1,641
HA_train	1,908
HE_train	1,011
HI_train	4,869
HS_train	809
HT_train	11
HU_train	39,270
HY_train	9,302
IA_train	4,909
ID_train	4,973
IG_train	9
IS_train	17
IT_train	172,828
JA_train	15,425
KA_train	215,015
KK_train	605
KL_train	11,064
KO_train	519
KY_train	1,790
LG_train	64,144
LI_train	2,304
LO_train	98
LT_train	12,895
LU_train	4,498
LV_train	4,410
MD_train	175
MH_train	186,565
MK_train	2,049
ML_train	1,235
MN_train	2,193
MR_train	16,514
MT_train	1,910
MY_train	1,241
NA_train	11,608
NB_train	227
NE_train	353
NH_train	23
NL_train	43,458
NN_train	464
NS_train	2
OC_train	304
OR_train	2,151
OS_train	414
PA_train	800
PL_train	24,173
PS_train	4,611
PT_train	22,923
QU_train	26
RM_train	2,148
RO_train	5,178
RU_train	26,654
RW_train	1,003,029
SA_train	2,528
SC_train	925
SD_train	271
SK_train	8,910
SL_train	1,469
SQ_train	2,658
SR_train	2,336
SV_train	8,150
SW_train	46,534
TA_train	46,390
TE_train	69
TG_train	123
TH_train	32,959
TI_train	2,010
TK_train	741
TN_train	1,078
TO_train	2,630
TR_train	40,377
TT_train	8,871
TW_train	205
UG_train	107,646
UK_train	26,773
UR_train	7,326
UZ_train	48,733
VI_train	2,104
VO_train	96
XH_train	7
YI_train	320
YO_train	1,404
YU_train	7,419
ZG_train	842
ZH_train	45,246
ZU_train	12
ZZ_train	734

Metadata Fields

Each metadata.json contains:

• dataset: Source dataset name
• language: Language code
• split: Data split (train/dev/test)
• sample_id: Original sample identifier
• text: Transcript
• duration_seconds: Audio duration in seconds
• chars_per_second: Text characters per second of audio
• original_sample_rate: Original audio sample rate
• dacvae_sample_rate: 48000 (DAC VAE input rate)
• latent_frames: Number of latent time frames
• Plus all original dataset-specific fields

---

Generated with Claude Code