---
language:
- en
license:
- cc-by-4.0
multilinguality:
- monolingual
size_categories:
- 10M<n<100M
source_datasets:
- amphion/Emilia-YODAS
pretty_name: NeuCodec Emilia-YODAS
tags:
- speech
- codec
- audio
- FSQ
- audio-compression
dataset_info:
  features:
  - name: _id
    dtype: string
  - name: dnsmos
    dtype: float64
  - name: duration
    dtype: float64
  - name: language
    dtype: string
  - name: phone_count
    dtype: int64
  - name: speaker
    dtype: string
  - name: text
    dtype: string
  - name: codes
    sequence: int64
  splits:
  - name: train
    num_bytes: 120073409108
    num_examples: 30592309
  download_size: 41111605599
  dataset_size: 120073409108
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# Dataset Card for NeuCodec Emilia-YODAS

## Dataset Description

- **Repository:** [NeuCodec repository](https://github.com/neuphonic/neucodec)
- **Paper:** Coming Soon

### Dataset Summary

The NeuCodec Emilia-YODAS dataset is an English-language dataset containing >30M audio samples (>78k hours), taken from the English-language subset of Emilia-YODAS and compressed with [NeuCodec](https://huggingface.co/neuphonic/neucodec).

# Usage

```python
import torch
from datasets import load_dataset
from neucodec import NeuCodec

# load dataset and model
dataset = load_dataset("neuphonic/emilia-yodas-english-neucodec", split="train", streaming=True)
model = NeuCodec.from_pretrained("neuphonic/neucodec")
model.eval()

# reconstruct a sample
fsq_codes = torch.tensor(next(iter(dataset))["codes"])[None, None, :]
print(f"FSQ codes shape: {fsq_codes.shape}")
recon = model.decode_code(fsq_codes)
print(f"Recon shape: {recon.shape}")
```

## Dataset Structure

### Data Instances

For each instance, the corresponding information from the Emilia-YODAS dataset is preserved. There is a column for id, dnsmos, duration, phone_count, speaker, text, and the codes.

```
{'id': 'EN_9eylmAUb-SQ_W000139',
 'dnsmos': 3.0064,
 'duration': 5.67,
 'phone_count': 102,
 'speaker': 'EN_9eylmAUb-SQ_SPEAKER_00',
 'text': 'In the Soviet Union during the 1920s, Yiddish...',
 'codes': [3254, 49895, 26484, 869, 23077, 27555, 20391,...]}
```

Each parquet file contains approx 200 MB. There are 241 parquet files. Each code sequence is meant to be used with our NeuCodec decoder, which currently supports an output sampling rate of 24 kHz.

### Data Fields

- `id`: a string containing the corresponding id from Emilia-YODAS
- `dnsmos`: a float containing the DNSMOS score from Emilia-YODAS
- `duration`: a float containing the sample duration from Emilia-YODAS
- `phone_count`: an integer of the number of phones in a sample from Emilia-YODAS
- `speaker`: a string containing the corresponding speaker from Emilia-YODAS
- `text`: a string containing the utterance text 
- `codes`: a list containing the compressed audio segment as NeuCodec codes


## Dataset Creation

### Source Data

The data was sourced from [Emilia-YODAS](https://huggingface.co/datasets/amphion/Emilia-Dataset)

## Additional Information

### Licensing Information

The NeuCodec Emilia-YODAS dataset is released under the [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/deed.en) license. 

### Citation Information

Coming soon