File size: 3,097 Bytes
7cdd294 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 | ---
license: other
license_name: see-source-datasets
tags:
- audio
- tts
- latents
- dac-vae
- speech
size_categories:
- 100K<n<1M
language:
- en
---
# EnvTTS Phase 1 — Pre-encoded DAC-VAE Latents
Pre-encoded audio latents for Phase 1 training of **EnvAudioEdit** (Small ~195 M CFM-DiT TTS model).
Audio from three English speech datasets is encoded offline with DAC-VAE (48 kHz, hop=1920),
saving GPU time during training by avoiding on-the-fly encoding.
---
## Contents
```
latents.zip
└── latents/
├── cv/ # 180 000 files (humanify/common_voice_english, 10%)
├── ps/ # 216 000 files (humanify/ps, 10%)
└── ht2/ # 313 000 files (humanify/ht2_44khz, 10%)
```
**Total: 709 000 `.pt` files, ~43 GB (zipped)**
---
## File Format
Each `.pt` file is a PyTorch tensor dict:
```python
{
"z": Tensor[T, 128], # DAC-VAE latent, float16
"text": str, # transcript
"length": int, # = T (number of latent frames)
}
```
| Field | Details |
|---|---|
| Audio codec | DAC-VAE (`matbee/sam-audio-small-onnx`) |
| Sample rate | 48 000 Hz |
| Hop length | 1 920 samples/frame |
| Latent dim | 128 |
| Max frames | 500 (≈ 20 s) — longer clips truncated |
| dtype | float16 |
Time ↔ frame conversion: `seconds = frames × 1920 / 48000`
---
## Source Datasets
| Subdir | Source | Size used |
|---|---|---|
| `cv` | [humanify/common_voice_english](https://huggingface.co/datasets/humanify/common_voice_english) | 10 % ≈ 180 K |
| `ps` | [humanify/ps](https://huggingface.co/datasets/humanify/ps) | 10 % ≈ 216 K |
| `ht2` | [humanify/ht2_44khz](https://huggingface.co/datasets/humanify/ht2_44khz) | 10 % ≈ 313 K |
> Original audio is licensed under the respective source dataset licenses.
> This dataset distributes only derived latent representations.
---
## Usage
### Extract
```bash
unzip latents.zip -d data/
# → data/latents/cv/, data/latents/ps/, data/latents/ht2/
```
### Load a single sample
```python
import torch
sample = torch.load("data/latents/cv/000000049.pt", weights_only=False)
z = sample["z"] # Tensor[T, 128], float16
text = sample["text"] # str
length = sample["length"] # int == z.shape[0]
```
### Use in EnvAudioEdit training (local latents mode)
Set `use_local_latents: true` in your training config and point `latent_dir` at the extracted directories:
```yaml
# configs/train_small_phase1.yaml
use_local_latents: true
latent_dir:
- "data/latents/cv"
- "data/latents/ps"
- "data/latents/ht2"
```
Then launch training:
```bash
accelerate launch scripts/train_phase1.py --config configs/train_small_phase1.yaml
```
---
## Encoding Environment
| Package | Version |
|---|---|
| onnxruntime-gpu | 1.23.2 |
| nvidia-cudnn-cu12 | 9.5.1.17 |
| torch | 2.11.0+cu130 |
> cuDNN 9.20 has a Conv1D bug; 8.x breaks PyTorch — pin to 9.5.1.17.
---
## Related
- Model checkpoints: [`<your-username>/envtts-small-phase1`](https://huggingface.co/<your-username>/envtts-small-phase1)
- Architecture & training plan: see the model repo README.
|