Japanese vowel native speech dataset

Infomation

Name: AIOI dataset
The AIOI dataset consists of 60 spoken sentences combined 5 words of 5 Japanese vowels, such as {aioi, aue, ao, ie, uo}. By connecting the words, 30 sentences that included all possible two-word sentences, e.g., “aioi ao,” “aue aue,” and “ie aioi,” and 5 three-word sentences, such as “ie ie uo,” “uo aue ie,” “ao ie ao,” “aue ao ie,” and “aioi uo ie” are prepared. Each sentence is spoken twice by a native Japanese speaker and recorded in the dataset.

File list

• DATA/ (Speech data.)
• HTKSCRIPT/ (Scripts to convert to MFCC using HTK.)
• NORMALIZE/ (Speech data which was adjusted volume and added a silent section in beginning and ending．)
• ORIGINAL/ (Original recorded data. No adjusted, e.g., sampling-rate.)
• PHONELABEL/ (Phoneme labels.)
• PHONELABELF/ (Phoneme labels in each frame.)
• WORDLABEL/ (Word labels.)
• WORDLAEELF/ (Word labels in each frame.)
• aioi_3dim/
  • DATA/ (3-dimensional MFCC features which compressed by deep sparse auto-encoder.)
  • LABEL/ (Label datas.)
• aioi_12dim/
  • DATA/ (12-dimensional MFCC features which picked from 39-dimentional MFCC features.)
  • LABEL/ (Label datas.)

Quick start

You can use 3 or 12 features in aioi_3dim or aioi_12dim directories. If you want to read features to Python code, you could include read using loadtxt function which in Numpy as follows.

import numpy as np
features = np.loadtxt("aioi_3dim/DATA/aioi_aioi.txt")

or

import numpy as np
all_features = np.load("aioi_3dim/npz/data.npz")
features = all_features["aioi_aioi"]

How to use

Preparation: install HTK.

1. Make a directory.

$ mkdir work

2. Copy files from DATA and HTKSCRIPT directories to work directory.

$ cp DATA/* work/
$ cp HTKSCRIPT/* work/

3. Move to work directory and run "htk.sh".

$ cd work
$ sh htk.sh

That's all!! The 39-dimensional MFCC features created in work directory as the text file.