---
block_size: 2048
sample_rate: 48000
latent_size: 11
vocoder: "046-multivoice-2048-48k-vlobeta-specdis-noise_824a15d4dc_streaming_norm.ts"
dataset: "VCTK"
vocoder_type: "RAVE"
alignment_type: "DCA"
likelihood_type: "NSF"
text_encoder_type: "CANINE"
---

# tungnaa_119_vctk

### dimensions

block size: 2048

sample rate: 48000

latent size: 11

### dataset

VCTK

### vocoder

`models/vocoder/046-multivoice-2048-48k-vlobeta-specdis-noise_824a15d4dc_streaming_norm.ts`


### training

```bash
tungnaa prep --datasets '{kind:"vctk", path:"/data/datasets/VCTK"}' --rave-path /data/users/victor/rave-v2/runs/046-multivoice-2048-48k-vlobeta-specdis-noise_824a15d4dc/version_0/checkpoints/046-multivoice-2048-48k-vlobeta-specdis-noise_824a15d4dc_streaming_norm.ts --out-path /data/users/victor/tmp/ivoice_prep_824a/

tungnaa trainer --experiment 119-vctk --model-dir /data/users/victor/ivoice-models --log-dir /data/users/victor/ivoice-logs --manifest /data/users/victor/tmp/ivoice_prep_824a/vctk.json --concat-speakers 2 --speaker-annotate --device cuda:1 --batch-size 32 --rave-model /data/users/victor/rave-v2/runs/046-multivoice-2048-48k-vlobeta-specdis-noise_824a15d4dc/version_0/checkpoints/046-multivoice-2048-48k-vlobeta-specdis-noise_824a15d4dc_streaming_norm.ts --lr 3e-4 --lr-text 3e-5 --epoch-size 200 --save-epochs 20 train
```

### notes

trained with concatation of utterance pairs plus speaker annotations. example syntax: `[p225] this is an utterance. [p330] this is another.`

uses a multi-dataset vocoder which was *not* fine tuned to only VCTK, so it should have a lot of play in the latent biases.

this model uses a neural spline flow likelihood.