File size: 1,741 Bytes
fbf1b69
aa6d0a8
 
 
 
 
 
 
 
fbf1b69
aa6d0a8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
language: sw
license: cc-by-sa-4.0
tags:
  - audio
  - text-to-speech
inference: false
datasets:
  - bookbot/OpenBible_Swahili
---

# VITS Base sw-KE-OpenBible

VITS Base sw-KE-OpenBible is an end-to-end text-to-speech model based on the [VITS](https://arxiv.org/abs/2106.06103) architecture. This model was trained from scratch on a real audio dataset. The list of real speakers include:

- sw-KE-OpenBible

The model's [vocabulary](https://huggingface.co/bookbot/vits-base-sw-KE-OpenBible/blob/main/symbols.py) contains the different IPA phonemes found in [gruut](https://github.com/rhasspy/gruut).

This model was trained using [VITS](https://github.com/jaywalnut310/vits) framework. All training was done on a Scaleway L40S VM with a NVIDIA L40S GPU. All necessary scripts used for training could be found in the [Files and versions](https://huggingface.co/bookbot/vits-base-sw-KE-OpenBible/tree/main) tab, as well as the [Training metrics](https://huggingface.co/bookbot/vits-base-sw-KE-OpenBible/tensorboard) logged via Tensorboard.

## Model

| Model                     | SR (Hz) | Mel range (Hz) | FFT / Hop / Win   | #epochs |
| ------------------------- | ------- | -------------- | ----------------- | ------- |
| VITS Base sw-KE-OpenBible | 44.1K   | 0-null         | 2048 / 512 / 2048 | 12000   |

## Training procedure

### Prepare Data

```sh
python preprocess.py \
    --text_index 1 \
    --filelists filelists/sw-KE-OpenBible_text_train_filelist.txt filelists/sw-KE-OpenBible_text_val_filelist.txt \
    --text_cleaners swahili_cleaners
```

### Train

```sh
python train.py -c configs/sw_ke_openbible_base.json -m sw_ke_openbible_base
```

## Frameworks

- PyTorch 2.2.2
- [VITS](https://github.com/bookbot-hive/vits)