---
license: apache-2.0
task_categories:
- text-to-speech
language:
- en
tags:
- kokoro-82M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: voices
    dtype: string
  - name: tensor
    sequence:
      sequence:
        sequence: float64
  splits:
  - name: train
    num_bytes: 11556949
    num_examples: 11
  download_size: 6344291
  dataset_size: 11556949
---

# Kokoro-82M Voices

This dataset contains all the voices available in [hexgrad/Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M).
This dataset provides the voices in 3 different formats.
1. Individual voices embeddings in different JSON file
2. Single JSON which contains all the voices in a JSON object.
3. Parquet format for usage via `datasets`
The voices name is the same as the `.pth` file names shown below.

```python
voices = [
    "af",
    "af_bella",
    "af_nicole",
    "af_sarah",
    "af_sky",
    "am_adam",
    "am_michael",
    "bf_emma",
    "bf_isabella",
    "bm_george",
    "bm_lewis",
]
```

For example, the "af" voices embeddings can be obtained using the "af.json" file which contains the embeddings as a JSON array.
The same "af" voices can be obtain in the single file by accessing the "af" name in the JSON object.
See `Usage` for more information on how to retreive the embeddings.

# Usage

Some example usage of downloading the "af" voice embedding.

## Downloading the JSON All

```python
# pip install -U -q huggingface_hub numpy
import json

import numpy as np
from huggingface_hub import hf_hub_download

file = hf_hub_download(
    repo_id="ecyht2/kokoro-82M-voices",
    repo_type="dataset",
    filename="voices.json"
)
with open(file, "r", encoding="utf-8") as f:
    voices = json.load(f)
    voice = np.array(voices.get("af")) # Change "af" to the voice that you want
print(voice.shape)
# (511, 1, 256)
```

## Downloading the JSON Single

```python
# pip install -U -q huggingface_hub numpy
import json

import numpy as np
from huggingface_hub import hf_hub_download

file = hf_hub_download(
    repo_id="ecyht2/kokoro-82M-voices",
    repo_type="dataset",
    filename="af.json" # Change "af" to the voice that you want
)
with open(file, "r", encoding="utf-8") as f:
    voice = json.load(f)
    voice = np.array(voice)
print(voice.shape)
# (511, 1, 256)
```

## Using Huggingface Datasets

```python
# pip install -U -q datasets numpy
from datasets import load_dataset
ds = load_dataset(repo, split="train")

for i in ds:
    if i["voices"] == "af":
        voice = np.array(i["tensor"])
        break
print(voice.shape)
# (511, 1, 256)
```