datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
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:… See the full description on the dataset page: https://huggingface.co/datasets/TTS-AGI/commonvoice22-sidon-dacvae.dacvae-tts-tr-w512-clean
DACVAE-TTS Turkish run C (width 512, clean data)
Generated audio of every evaluated checkpoint of the training run tr-w512-clean (Turkish zero-shot voice-cloning TTS,
dacvae-tts, frozen Meta DACVAE latents, 48 kHz). This repository holds
model outputs and metrics, not training data. Training data: Vyvo/tr-dataset-12 (Turkish podcast segments).
Run C: configs/nano_tr_w512.yaml (66.5M parameters: width 512, 8 heads, batch expansion 4, frame budget 6000), trained from scratch on… See the full description on the dataset page: https://huggingface.co/datasets/VoiceHub/dacvae-tts-tr-w512-clean.dacvae-tts-tr-trc-w640
dacvae-tts Turkish 115M: trc-w640 (tr-combined + tr-dataset-12)
Checkpoints of the training run trc-w640, each with 5 outputs: Turkish zero-shot voice-cloning TTS
(dacvae-tts, branch turkish-tts; flow-matching DiT on frozen Meta DACVAE
latents, 48 kHz). Training in progress (update 61200 of 100000): every 10000 updates a new checkpoint and its 5 outputs are added here.
5 outputs per checkpoint
The same 5 unseen sentences, each spoken in the voice of a different… See the full description on the dataset page: https://huggingface.co/datasets/VoiceHub/dacvae-tts-tr-trc-w640.dacvae-tts-tr-nano-b-ke4
DACVAE-TTS Turkish run B (batch expansion 4)
Generated audio of every evaluated checkpoint of the training run tr-nano-b-ke4 (Turkish zero-shot voice-cloning TTS,
dacvae-tts, frozen Meta DACVAE latents, 48 kHz). This repository holds
model outputs and metrics, not training data. Training data: Vyvo/tr-dataset-12 (Turkish podcast segments).
Run B: configs/nano_tr_ke4.yaml (same 51.4M model as run A, context-sharing batch expansion 4, frame budget 7000), same data (~77 h), one RTX… See the full description on the dataset page: https://huggingface.co/datasets/VoiceHub/dacvae-tts-tr-nano-b-ke4.dacvae-tts-tr-nano-a
DACVAE-TTS Turkish run A (nano recipe, full data)
Generated audio of every evaluated checkpoint of the training run tr-nano-a (Turkish zero-shot voice-cloning TTS,
dacvae-tts, frozen Meta DACVAE latents, 48 kHz). This repository holds
model outputs and metrics, not training data. Training data: Vyvo/tr-dataset-12 (Turkish podcast segments).
Run A: configs/nano_tr.yaml (51.4M parameters), 17 shards of Vyvo/tr-dataset-12 with quality >= 55 (~77 h train), one RTX 4090, frame budget… See the full description on the dataset page: https://huggingface.co/datasets/VoiceHub/dacvae-tts-tr-nano-a.dacvae-tts-tr-stage2-b
DACVAE-TTS Turkish stage 2 (from run B 40k, clean data)
Generated audio of every evaluated checkpoint of the training run tr-stage2-b (Turkish zero-shot voice-cloning TTS,
dacvae-tts, frozen Meta DACVAE latents, 48 kHz). This repository holds
model outputs and metrics, not training data. Training data: Vyvo/tr-dataset-12 (Turkish podcast segments).
Stage 2: warm start (--init-from) from run B's 40k checkpoint (VoiceHub/dacvae-tts-tr-nano-b-ke4), trained 20k more updates on the… See the full description on the dataset page: https://huggingface.co/datasets/VoiceHub/dacvae-tts-tr-stage2-b.mls-enhanced-dacvae
Multilingual LibriSpeech converted to DAC VAE latents
Source
facebook/multilingual_librispeech
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:… See the full description on the dataset page: https://huggingface.co/datasets/TTS-AGI/mls-enhanced-dacvae.dacvae-tts-tr-w512-stage2-hq
DACVAE-TTS Turkish run C stage 2 (width 512, high-quality subset)
Generated audio of every evaluated checkpoint of the training run tr-w512-stage2-hq (Turkish zero-shot voice-cloning TTS,
dacvae-tts, frozen Meta DACVAE latents, 48 kHz). This repository holds
model outputs and metrics, not training data. Training data: Vyvo/tr-dataset-12 (Turkish podcast segments).
Warm start from VoiceHub/dacvae-tts-tr-w512-clean (step 60k, 66.5M parameters), 10k more updates on the high-quality… See the full description on the dataset page: https://huggingface.co/datasets/VoiceHub/dacvae-tts-tr-w512-stage2-hq.vocal-bursts-taxonomy-DACVAE
Vocal Bursts Taxonomy — DACVAE + MaestroClap Embeddings & Scores
Processed version of with DACVAE latents, MaestroClap embeddings, derived attribute/quality/speaker scores, and Gemini-verified labels.
Overview
Metric
Value
Total samples
16,175
Categories
82
Genders
male, female
Female samples
8,097
Male samples
8,078
Gemini Label Verification
Every sample was sent to Gemini 3.1 Flash Lite for two independent tasks:
Match scoring:… See the full description on the dataset page: https://huggingface.co/datasets/TTS-AGI/vocal-bursts-taxonomy-DACVAE.Emotion-Voice-Attribute-Reference-Snippets-DACVAE-Wave
Emotion and Voice Attribute Reference Snippets - DACVAE and Wave
Merged dataset combining TTS-AGI/enhanced-emo-snippets-balanced-DACVAE and
TTS-AGI/emotion-attribute-conditioning-dacvae with decoded WAV audio.
Overview
Total samples: 606,178
Filtered out: 363,331 (samples with speech_quality < 1.8)
Total tar files: 328
Total size: 1.54 TB
Audio format: WAV, 48kHz, PCM 16-bit mono
Latents: DAC-VAE float16 [T, 128] at 25 frames/sec
Dimensions: 57 (40 emotions + 15 voice… See the full description on the dataset page: https://huggingface.co/datasets/TTS-AGI/Emotion-Voice-Attribute-Reference-Snippets-DACVAE-Wave.podcast-dramabox-dacvae-pairs
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… See the full description on the dataset page: https://huggingface.co/datasets/TTS-AGI/podcast-dramabox-dacvae-pairs.dacvae-tts-tr-stage3-hq
tr-stage3-hq
Generated audio of every evaluated checkpoint of the training run tr-stage3-hq (Turkish zero-shot voice-cloning TTS,
dacvae-tts, frozen Meta DACVAE latents, 48 kHz). This repository holds
model outputs and metrics, not training data. Training data: Vyvo/tr-dataset-12 (Turkish podcast segments).
How to listen
monitor/prompt-<uid>.wav: the reference voice given to the model (a real validation recording of an unseen speaker,
decoded through the DACVAE… See the full description on the dataset page: https://huggingface.co/datasets/VoiceHub/dacvae-tts-tr-stage3-hq.dacvae-tts-tr-pilot
DACVAE-TTS Turkish pilot (4 shards, 4k updates)
Generated audio of every evaluated checkpoint of the training run tr-pilot (Turkish zero-shot voice-cloning TTS,
dacvae-tts, frozen Meta DACVAE latents, 48 kHz). This repository holds
model outputs and metrics, not training data. Training data: Vyvo/tr-dataset-12 (Turkish podcast segments).
Pilot run on 4 of 17 shards (~21 h) to validate the Turkish pipeline; stopped at 4k updates. extra/ceiling-pilot holds real validation speech… See the full description on the dataset page: https://huggingface.co/datasets/VoiceHub/dacvae-tts-tr-pilot.mls-dacvae
Multilingual LibriSpeech converted to DAC VAE latents
Source
facebook/multilingual_librispeech
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:… See the full description on the dataset page: https://huggingface.co/datasets/TTS-AGI/mls-dacvae.Emotion-Voice-Attribute-Reference-Snippets-DACVAE
Emotion and Voice Attribute Reference Snippets - DACVAE and Wave
Merged dataset combining TTS-AGI/enhanced-emo-snippets-balanced-DACVAE and
TTS-AGI/emotion-attribute-conditioning-dacvae with decoded WAV audio.
Overview
Total samples: 606,178
Filtered out: 363,331 (samples with speech_quality < 1.8)
Total tar files: 328
Total size: ~98 GB (latents-only, no WAV)
Audio format: WAV, 48kHz, PCM 16-bit mono
Latents: DAC-VAE float16 [T, 128] at 25 frames/sec
Dimensions: 57 (40… See the full description on the dataset page: https://huggingface.co/datasets/TTS-AGI/Emotion-Voice-Attribute-Reference-Snippets-DACVAE.eurospeech-enhanced-dacvae
EuroSpeech parliamentary speech converted to DAC VAE latents
Source
disco-eth/EuroSpeech
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:… See the full description on the dataset page: https://huggingface.co/datasets/laion/eurospeech-enhanced-dacvae.