The complete working backup of the data behind DACFlow-EN-10k, an English text-to-speech model (VoiceHub/DACFlow-EN-10k). It holds everything needed to resume the training and the evaluations if the training machine is lost: the tokenized training data, the REPA targets, the quality annotations, the manifests and every catalog, the held-out voices' audio, the evaluation sets, the experiment data and the processing bookkeeping. Every file is checked against the Hub's own record of it.

Looking for the dataset? Start with VoiceHub/DACFlow-EN-10k-data: only the tokenized training data (the codec latents, their texts and the exact training list), with a guide to use it.

Want to listen to the model? Its samples are on VoiceHub/DACFlow-EN-10k.


What this repo is	the project's complete working backup (disaster recovery, and the record of every input); not a dataset release
The dataset	VoiceHub/DACFlow-EN-10k-data (the tokenized training data only)
The model, its checkpoints and samples	VoiceHub/DACFlow-EN-10k
The main run's training state (to resume it)	VoiceHub/mytts-en-base (prefix `en-base-s1`); not here
Made from	SynDataLab-EN/echo-clones-4m-en (revision `aebc7c787c64`)
Status	complete, updated 2026-10-03 00:22:12 UTC
Size now	36,896 files, 522.6 GB
Licence	apache-2.0 (see Provenance and licence)
Code	kadirnar/dacvae-next (private at the moment), Python package `mytts`

What is in it

What each part is and what it is needed for (Layout has one line per folder). The latents and the REPA targets come at 25 frames per second, so frame i of one lines up with frame i of the other.

part	in plain words	exact form	needed to	folder
Codec latents (what the model learns to generate)	a compressed form of each clip's audio, with its text and voice; the codec's decoder turns it back into sound	Semantic-DACVAE posterior mean, 25 Hz x 128 numbers, float16 (codec `Aratako/Semantic-DACVAE-Japanese@737fbad2a679f880`)	train (the same files are on VoiceHub/DACFlow-EN-10k-data)	`tts_en/latents/echo_en/`
Catalogs	the filtered clip lists a run trains on; `echo_en_v1` is the main run's	`<name>/catalog.parquet` + `version.json`	train	`tts_en/catalogs/`
Latent statistics	the frozen normalization of the latents and the latent of silence	`latent_stats.json`, `silence_latent.npy`	train, synthesize	`tts_en/`
REPA targets (a training aid)	what a large speech-understanding model hears in each frame; the TTS model is nudged to match it, which makes it learn faster	w2v-BERT 2.0 layer 16, reduced from 1,024 to 128 numbers with PCA (`tts_en/ssl/pca.npz`), float16	resume the training (its REPA loss)	`tts_en/ssl/`
Quality annotations	what a speech recognizer heard, where the silences are; used to drop bad clips	Parakeet CTC 1.1B transcript + CER / WER / NLL of every clip; Whisper-large-v3-turbo on a 2 % sample; silence at the edges, the longest pause and whether the clip hit EchoTTS's length cap	rebuild or audit a catalog	`tts_en/annotations/`
Manifests	the source row of every clip	one JSON line per clip	provenance	`tts_en/manifests/`
Held-out voices	236 voices never trained on: their clips' latents, REPA targets, annotations and original audio	as above, plus `.wav`	evaluation only	`*/echo_en_heldout/`, `eval/en/echo_heldout_audio/`
Evaluation sets	the echo tests, Seed-TTS test-en, LibriSpeech(-PC), AMI, validation runs, the protect list	lists + audio	evaluation	`eval/en/`
Experiments	HuBERT REPA targets of the screening subset (EN-25 teacher screen)	as the REPA targets	not used by the main run	`tts_en_hubl*/`
Bookkeeping	the processing state and ledger; this backup's file list, status and symlinks	JSON / JSONL	resume the processing; verify and restore this backup	`tts_en/ingest/`, `backup/`

What it is not

Not the place to start. To use the data, start with VoiceHub/DACFlow-EN-10k-data: the tokenized training data alone, with a guide. This repo holds much more, because it must be able to bring back everything.
Not a place to listen. Latents are numbers. The model's samples are on VoiceHub/DACFlow-EN-10k. The only audio here is evaluation material: the held-out voices' clips (eval/en/echo_heldout_audio/) and the evaluation sets.
Not the source dataset. The original audio and texts are SynDataLab-EN/echo-clones-4m-en; every clip here points back to its source row (tts_en/manifests/, and source / id in each .meta.parquet).
Not usable with another codec. The latents belong to this exact codec version; the REPA targets to tts_en/ssl/pca.npz.
No real recordings in the training data. All of the training speech is synthetic (EchoTTS voice clones). Only the evaluation sets under eval/en/ hold real recordings (LibriSpeech, AMI, the Common Voice prompts of Seed-TTS-eval).

Size and progress

	now	when complete
Source files (parquet) processed	7,967 / 7,967 (100.0 %)	7,967
Source clips read	3,983,250	about 3.98 M
Groups uploaded (16 source files each)	499	498
Held-out clips kept (236 voices, at most 60 each)	14,160	at most 14,160
Files in this repo	36,896	-
Size	522.6 GB	522.6 GB

The stream is complete; this page is written again whenever files are added. Last update: 2026-10-03 00:22:12 UTC (--force).

Layout

folder	what is inside	used for	files	size
`tts_en/latents/echo_en/`	codec latents of every training clip: `shard-<tag>-NNNNNN.bin` (float16, 25 frames/s x 128 numbers), `.idx.json` (where each clip starts and how long it is), `.meta.parquet` (text, voice, duration, source row of each clip)	training	1,497	250.3 GB
`tts_en/ssl/echo_en/`	REPA targets of every training clip: w2v-BERT 2.0 layer 16, PCA to 128 numbers, float16, one row per latent frame (`shard-<tag>-t0-NNNNNN.{bin,idx.json}`)	training	998	249.8 GB
`tts_en/ssl/`	`pca.npz`: the PCA that turns w2v-BERT's 1,024 numbers into 128; `info.json`	training	2	529.8 kB
`tts_en/annotations/asr2/echo_en/`	Parakeet CTC 1.1B transcript of every clip with CER / WER / NLL against its text (`part-<tag>.parquet`)	training (clip filter)	499	457.4 MB
`tts_en/annotations/asr/echo_en/`	Whisper-large-v3-turbo transcript of a 2 % sample: a cross-check of the Parakeet scores	analysis	499	15.2 MB
`tts_en/annotations/edges/echo_en/`	silence at the start and the end of every clip	training (clip edges)	499	196.8 MB
`tts_en/annotations/silence/echo_en/`	longest pause inside every clip, the generation length cap and whether the clip reached it (EchoTTS failure clips)	training (clip filter)	499	105.4 MB
`tts_en/manifests/`	one JSON line per clip: id, text, voice, licence and its source row `hf://datasets/<source>@<revision>/<file>#<row>` (held-out clips: their kept audio file)	provenance	499	1.5 GB
`tts_en/catalogs/`	catalogs: the filtered clip lists a run trains on (`<name>/catalog.parquet` + `version.json`, build log, leak check)	training	12	496.3 MB
`tts_en/`	`latent_stats.json` (the frozen latent normalization of the model), `splits.json`, file lists and held-out voice lists	training	8	214.3 kB
`tts_en/ingest/`	the ingest's state (run spec, finished source files) and ledger (one line per group), as of the last upload	bookkeeping	2	1.8 MB
`tts_en/manifests_local/`	the Tier-1 screening subset's manifest (local raw paths; provenance only)	bookkeeping	1	12.9 MB
`tts_en/latents/echo_en_heldout/`	codec latents of the held-out voices' clips	held out: evaluation only	111	946.5 MB
`tts_en/ssl/echo_en_heldout/`	REPA targets of the held-out voices' clips	held out: evaluation only	74	944.3 MB
`tts_en/annotations/*/echo_en_heldout/`	the four annotations of the held-out clips (Whisper on every one)	held out: evaluation only	148	5.0 MB
`eval/en/echo_heldout_audio/`	the original audio of up to 60 clips per held-out voice (`<voice>/<id>.wav`): the prompts and references of the echo tests	held out: evaluation only	14,160	5.5 GB
`eval/en/echo_heldout/`	the echo tests of the held-out voices: echo-dev / echo-test lists, prompt draws, prompt pairs (with their prompt audio)	evaluation only	914	542.5 MB
`eval/en/echo_heldout_ann/`	annotations and manifests of the kept held-out audio	evaluation only	7	1.6 MB
`eval/en/seedtts_testset/`	Seed-TTS-eval test sets (`en`: Common Voice prompts; `zh`), CC BY 4.0	evaluation only	5,134	1.2 GB
`eval/en/LibriSpeech/`	LibriSpeech test-clean / dev-clean audio (the LibriSpeech-PC test), CC BY 4.0	evaluation only	5,512	727.8 MB
`eval/en/ami/`	AMI Meeting Corpus clips (`ihm` headset, `sdm` distant microphone), CC BY 4.0	evaluation only	996	336.5 MB
`eval/en/validation/`	validation gates: evaluation runs (scores and audio) of the real test audio and of its codec resynthesis; they check the evaluation pipeline	evaluation only	4,732	1.8 GB
`eval/en/voice_emb/`	speaker embeddings of eval prompts and references	evaluation only	8	83.2 MB
`eval/en/`	evaluation lists (`*.jsonl`: Seed-TTS test-en dev / test, LibriSpeech-PC, AMI) and the protect list (test sentences no training catalog may hold)	evaluation only	18	15.1 MB
`tts_en_hubl18/`	HuBERT-large layer 18 REPA targets of the screening subset (EN-25 teacher screen)	experiment (not used by the main run)	25	2.5 GB
`tts_en_hubl24/`	HuBERT-large layer 24 REPA targets of the screening subset (EN-25)	experiment (not used by the main run)	25	2.5 GB
`tts_en_hubl24_fp32/`	the same, computed in fp32 (EN-25)	experiment (not used by the main run)	17	2.5 GB
`backup/`	`manifest.jsonl` (every file: size, sha256, group), `status.json` (the numbers on this page), `symlinks.json`	bookkeeping	3	-

<tag> = g<time_ns>-<group> names one ingest group (16 source files, about 8,000 clips). Every file of a group carries its tag, so a group can be downloaded on its own. echo_en is the training source, echo_en_heldout the held-out voices.

Held-out and evaluation-only parts

Do not train on these if you want results that compare with ours:

Held-out voices: every echo_en_heldout folder and eval/en/echo_heldout_audio/. 236 of the 4,000 voices (list) never enter training; at most 60 clips of each are kept for the echo tests (EN-08).
Evaluation sets: everything under eval/en/. Seed-TTS test-en, LibriSpeech(-PC) and AMI come from other corpora and keep their own licences.
Echo-unseen voices: 177 more voices (list, EN-56) are inside echo_en but no training catalog holds them (the catalogs already leave them out, and the trainer refuses a catalog that does not).

Backup and restore (maintainers)

scripts/hub_backup.py uploads only what the ingest has committed, checks every file after its commit (size, plus the LFS sha256 or, for small files, the git blob sha1) and, before every status update, that every file of backup/manifest.jsonl is on the repo with that size and hash. The LFS sha256 is the Hub's record of the upload, not a re-read of the stored bytes: restore hashes what it downloads.

python scripts/hub_backup.py restore --repo VoiceHub/DACFlow-EN-10k-backup --out /workspace/restore_check --sample-groups 3   # a verified sample
python scripts/hub_backup.py restore --repo VoiceHub/DACFlow-EN-10k-backup --all --out /workspace/data_restore              # everything, verified
# then, with the ingest and its job_guard stopped and any existing /workspace/data/<root> moved aside:
for r in tts_en tts_en_hubl18 tts_en_hubl24 tts_en_hubl24_fp32; do mv -T /workspace/data_restore/$r /workspace/data/$r; done
mkdir -p /workspace/data/eval && mv -T /workspace/data_restore/eval/en /workspace/data/eval/en

restore writes only into an empty directory (--into-empty-only / --overwrite-existing allow a non-empty one) and always refuses the data root of a running ingest. The manifests and the ingest state record absolute paths under /workspace/data, so they hold again after the swap, and the ingest resumes from tts_en/ingest/echo_en.json. The main run's checkpoints are on VoiceHub/DACFlow-EN-10k, its training state (optimizer included) in VoiceHub/mytts-en-base (prefix en-base-s1): neither is here.

Use a part of it

For the training data alone, VoiceHub/DACFlow-EN-10k-data is simpler (same latents, with a numpy reader and a decoder snippet).

Download

All of it is large (522.6 GB now). One group with its REPA targets:

from huggingface_hub import snapshot_download

tag = "g1790656525369654176-00001"  # any group; backup/status.json lists them all
local = snapshot_download("VoiceHub/DACFlow-EN-10k-backup", repo_type="dataset", local_dir="DACFlow-EN-10k-backup",
                          allow_patterns=[f"tts_en/*/echo_en/*{tag}*", "tts_en/ssl/pca.npz", "tts_en/latent_stats.json"])

With the project code

from mytts.config import DataConfig
from mytts.data.catalog import CatalogFilters, build_catalog
from mytts.data.dataset import Catalog, TTSDataset

root = f"{local}/tts_en"
cat_dir, version = build_catalog(root, "my_subset", CatalogFilters(), num_workers=2, verbose=False)  # index what you downloaded
cat = Catalog(cat_dir, split=None)
ds = TTSDataset(cat, root, DataConfig(use_ssl=True), latent_dim=128, ssl_dim=(version.filters or {}).get("ssl_dim"), train=False)
item = ds[0]
print(cat.get(0, "text"), item["latent"].shape, item["ssl"].shape)  # [frames, 128] each

The catalog: the exact training list

tts_en/catalogs/<name>/catalog.parquet has one row per clip that passed the quality filters: key, text, voice, length, where its latents and REPA targets are (paths relative to tts_en/) and its split (train / val); version.json records the filters and the hours. echo_en_v1 is the catalog of the main run (here): Parakeet CER <= 0.3 (rescored with the en-v2 text normalization), no long pause (> 6 s), no clip stuck at the length cap with silence, the echo-unseen voices left out, 0.1 % held back as val. en_t1 is the 88.8 h screening subset. Catalogs here: echo_en_v1, en_t1, en_t1_x56. To train on exactly what the model saw, download all of tts_en/, then Catalog(f"{root}/catalogs/echo_en_v1", split="train").

How it was made

Stream (scripts/ingest_stream.py, EN-12, issue #36): the source's 7,967 parquet files (1.51 TB) are downloaded 16 at a time (one group), in a fixed shuffled order.
Decode each clip once. Clips of the 236 held-out voices go to echo_en_heldout (the first 60 of each voice, with their audio kept); all others to echo_en.
Encode on the GPU: the codec latents (48 kHz audio, exact fp32 encoder, stored as float16) and the w2v-BERT 2.0 layer-16 features, projected to 128 numbers with a PCA fitted once (the REPA targets).
Annotate: Parakeet CTC transcript and CER / WER / NLL for every clip, Whisper on the held-out clips and a 2 % sample, silence at the clip edges, the longest pause and the length cap.
Check and commit: a group counts only when every output covers its clips (the coverage gate); then it gets its line in the ledger (tts_en/ingest/) and its downloaded source files are deleted.
Back up (scripts/hub_backup.py): after every 50 committed groups the new files come here, each checked against the Hub's own record of it (size and sha256).
Curate (after the stream, scripts/recipes/en_m4.sh catalog): the filters above make the training catalog echo_en_v1.
Publish (scripts/hub_data_release.py): the training latents, the echo_en_v1 catalog and the latent statistics go to the data page VoiceHub/DACFlow-EN-10k-data.

Provenance and licence

This dataset: apache-2.0 (owner decision, 2026-09-29).
Source: SynDataLab-EN/echo-clones-4m-en; its card states apache-2.0. Its speech was generated with EchoTTS (jordand/echo-tts-base, cc-by-nc-sa-4.0 on its card) from the reference voices of SynDataLab-EN-Refs/echo-ref-speakers-4k-en and the texts of SynDataLab-EN-Refs/echo-4m-text-en (neither card states a licence). Check these terms yourself before any commercial use.
Models used to make the features (each under its own licence; none of their weights are in this repo): Aratako/Semantic-DACVAE-Japanese (MIT), facebook/w2v-bert-2.0 (MIT), nvidia/parakeet-ctc-1.1b (CC BY 4.0), openai/whisper-large-v3-turbo (MIT); facebook/hubert-large-ll60k (Apache-2.0) for the tts_en_hubl* roots only.
Evaluation audio under eval/en/ keeps its own licence and attribution: Seed-TTS-eval (CC BY 4.0, Common Voice prompts), LibriSpeech (CC BY 4.0), AMI Meeting Corpus (CC BY 4.0, University of Edinburgh / AMI consortium).

DACFlow-EN-10k-backup