Update README.md
Browse files
README.md
CHANGED
|
@@ -4,9 +4,9 @@ configs:
|
|
| 4 |
- config_name: 22khz
|
| 5 |
data_files:
|
| 6 |
- split: utterancces
|
| 7 |
-
path: "manifest_22khz.json"
|
| 8 |
- split: chapters
|
| 9 |
-
path: "chapters_22khz.json"
|
| 10 |
---
|
| 11 |
|
| 12 |
# HiFiTTS-2: A Large-Scale High Bandwidth Speech Dataset
|
|
@@ -21,6 +21,14 @@ padding: 0;
|
|
| 21 |
|
| 22 |
## Dataset Description
|
| 23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
## Download Instructions
|
| 25 |
|
| 26 |
1. Download the *manifet.json* file and *chapter.json* files corresponding to your desired sampling rate from this Hugging Face repository. Copy these into a workspace directory (in this example */home/hifitts2*).
|
|
@@ -50,8 +58,6 @@ python /home/NeMo-speech-data-processor/main.py \
|
|
| 50 |
max_workers=8
|
| 51 |
```
|
| 52 |
|
| 53 |
-
Due to legal and ethical considerations, we do not redistribute LibriVox audiobooks to download from Hugging Face.
|
| 54 |
-
|
| 55 |
## Dataset Owner(s)
|
| 56 |
|
| 57 |
NVIDIA Corporation
|
|
|
|
| 4 |
- config_name: 22khz
|
| 5 |
data_files:
|
| 6 |
- split: utterancces
|
| 7 |
+
path: "22khz/manifest_22khz.json"
|
| 8 |
- split: chapters
|
| 9 |
+
path: "22khz/chapters_22khz.json"
|
| 10 |
---
|
| 11 |
|
| 12 |
# HiFiTTS-2: A Large-Scale High Bandwidth Speech Dataset
|
|
|
|
| 21 |
|
| 22 |
## Dataset Description
|
| 23 |
|
| 24 |
+
This repository contains the metadata for HiFiTTS-2, a large scale speech dataset derived from LibriVox audiobooks. For more details, please refer to our paper. [TODO: Add ArXiv link]
|
| 25 |
+
|
| 26 |
+
The dataset contains metadata for approximately 36.7k hours of audio that can be downloaded from LibriVox at a 48 kHz sampling rate.
|
| 27 |
+
|
| 28 |
+
The metadata contains estimated bandwidth, which can be used to infer the original sampling rate the audio was recorded at. The base dataset is filtered for a bandwidth appropriate for training speech models at *22 kHz*. We also provide a precomputed subset with 31.7k hours appropraite for *44 kHz* training. Users can modify the download script to use any sampling rate and bandwidth threshold which might be more appropriate for their work.
|
| 29 |
+
|
| 30 |
+
Due to legal and ethical considerations, we do not redistribute LibriVox audiobooks on Hugging Face. All audio in the dataset can be downloaded from LibriVox, following the instructions below.
|
| 31 |
+
|
| 32 |
## Download Instructions
|
| 33 |
|
| 34 |
1. Download the *manifet.json* file and *chapter.json* files corresponding to your desired sampling rate from this Hugging Face repository. Copy these into a workspace directory (in this example */home/hifitts2*).
|
|
|
|
| 58 |
max_workers=8
|
| 59 |
```
|
| 60 |
|
|
|
|
|
|
|
| 61 |
## Dataset Owner(s)
|
| 62 |
|
| 63 |
NVIDIA Corporation
|