mytts-en-v1-data

The prepared training data of the English text-to-speech model VoiceHub/mytts-en-v1. Every clip of SynDataLab-EN/echo-clones-4m-en (about 3.98 M synthetic English speech clips of 4,000 voices) turned into the compact numbers the model learns from, plus the checks used to pick the good clips.

Model trained on it	VoiceHub/mytts-en-v1
Made from	SynDataLab-EN/echo-clones-4m-en (revision aebc7c787c64)
Status	being processed: 5,211 of 7,967 source files (65.4 %), updated 2026-10-01 17:53:00 UTC
Size now	34,449 files, 347.7 GB (about 522.4 GB when complete)
Licence	apache-2.0 (see Provenance and licence)
Code	kadirnar/dacvae-next (private at the moment)

What is in it

For every clip the model needs three things. The first two come at 25 frames per second, so frame i of the latents lines up with frame i of the REPA targets:

feature	in plain words	exact form	folder
Codec latents (what the model learns to generate)	a compressed form of the audio; 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)	tts_en/latents/
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	tts_en/ssl/
Text and quality checks	the sentence each clip says, what a speech recognizer heard, where the silences are; used to drop bad clips	the source text (.meta.parquet, manifests); 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	tts_en/annotations/

Also here: the original audio of the 236 held-out voices and the English evaluation sets (both evaluation only, see below).

What it is not

• Not audio to listen to. Latents are numbers. To hear the speech: play the source clips in the source's dataset viewer, listen to the model's samples on VoiceHub/mytts-en-v1, or decode latents with the codec (Hear a clip). The only audio here is the held-out voices' clips (eval/en/echo_heldout_audio/: open a .wav to play it) 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 the training list by itself. It holds every clip. The clips a model trains on (after the quality filters) are listed in a catalog under tts_en/catalogs/ (The catalog).
• 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	5,211 / 7,967 (65.4 %)	7,967
Source clips read	2,605,250	about 3.98 M
Groups uploaded (16 source files each)	326	498
Held-out clips kept (236 voices, at most 60 each)	14,160	at most 14,160
Files in this repo	34,449	-
Size	347.7 GB	about 522.4 GB

The data is still being processed (EN-12); this page is written again after every 50 new groups. Last update: 2026-10-01 17:53:00 UTC (50 new committed groups (>= 50)).

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	978	163.7 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	652	163.3 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)	326	299.1 MB
tts_en/annotations/asr/echo_en/	Whisper-large-v3-turbo transcript of a 2 % sample: a cross-check of the Parakeet scores	analysis	326	11.4 MB
tts_en/annotations/edges/echo_en/	silence at the start and the end of every clip	training (clip edges)	326	128.7 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)	326	69.0 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	326	988.2 MB
tts_en/catalogs/	catalogs: the filtered clip lists a run trains on (<name>/catalog.parquet + version.json, build log, leak check)	training	8	9.8 MB
tts_en/	latent_stats.json (the frozen latent normalization of the model), splits.json, file lists and held-out voice lists	training	7	213.7 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	1.2 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	212	111.4 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	8	3.0 MB
tts_en_hubl18/	HuBERT-large layer 18 REPA targets of the screening subset (EN-25 teacher screen)	experiment (not used by v1)	25	2.5 GB
tts_en_hubl24/	HuBERT-large layer 24 REPA targets of the screening subset (EN-25)	experiment (not used by v1)	25	2.5 GB
tts_en_hubl24_fp32/	the same, computed in fp32 (EN-25)	experiment (not used by v1)	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).

How to load

Download

All of it is large (347.7 GB now). One group is enough to try the data:

from huggingface_hub import snapshot_download

tag = "g1790656525369654176-00001"  # any group; backup/status.json lists them all
local = snapshot_download("VoiceHub/mytts-en-v1-data", repo_type="dataset", local_dir="mytts-en-v1-data",
                          allow_patterns=[f"tts_en/*/echo_en/*{tag}*", "tts_en/ssl/pca.npz", "tts_en/latent_stats.json"])

Everything: allow_patterns=["tts_en/*"] (the training data) or no filter, or the verified restore below.

Read one clip (numpy only)

import json

import numpy as np
import pyarrow.parquet as pq

shard = f"{local}/tts_en/latents/echo_en/shard-{tag}-000000"
idx = json.load(open(shard + ".idx.json"))  # keys, offsets, n_frames, dim, dtype
meta = pq.read_table(shard + ".meta.parquet").to_pylist()  # one row per clip: key, text, speaker, duration, ...
lat = np.memmap(shard + ".bin", dtype=idx["dtype"], mode="r").reshape(-1, idx["dim"])
i = 0
z = np.asarray(lat[idx["offsets"][i] : idx["offsets"][i] + idx["n_frames"][i]], np.float32)  # [frames, 128], 25 frames per second
print(meta[i]["text"], z.shape)

The REPA targets of the same clip: tts_en/ssl/echo_en/shard-<tag>-t0-000000.{bin,idx.json}, same key, same frames.

With the project code (mytts)

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 (uploaded after the stream completes): 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: 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").

Hear a clip

import soundfile as sf
import torch

from mytts.codec import DACVAECodec  # needs pip install git+https://github.com/facebookresearch/dacvae

codec = DACVAECodec.load("Aratako/Semantic-DACVAE-Japanese@737fbad2a679f880")  # the codec the latents were made with
wav = codec.decode(torch.from_numpy(z)[None])[0].numpy()  # z from 'Read one clip'; 48 kHz
sf.write("clip.wav", wav, 48000)

How it was made

1. 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.
2. 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.
3. 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).
4. 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.
5. 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.
6. Upload (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).
7. Curate (after the stream, scripts/recipes/en_m4.sh catalog): the filters above make the training catalog echo_en_v1.

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 SynData-2/echo-ref-speakers-4k-en and the texts of SynData-2/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).

Backup and restore (maintainers)

This repo is also the project's backup of its data roots. 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/mytts-en-v1-data --out /workspace/restore_check --sample-groups 3   # a verified sample
python scripts/hub_backup.py restore --repo VoiceHub/mytts-en-v1-data --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.

Links

• Model: VoiceHub/mytts-en-v1
• Source audio and texts: SynDataLab-EN/echo-clones-4m-en
• Code: kadirnar/dacvae-next (branch roadmap/en-echo): scripts/ingest_stream.py, mytts/data/, scripts/hub_backup.py
• The project's earlier pages (screening phase, kept as an archive; not needed to use this data): VoiceHub/mytts-en (screening checkpoints), VoiceHub/mytts-en-samples (their samples), VoiceHub/mytts-en-ablations (screening runs and results), VoiceHub/mytts-en-eval (evaluation lists), VoiceHub/mytts-en-base (the main run's internal training state)