RAWDysPeech: A Preprocessed Raw Audio Dataset For Speech Dysarthria

RAWDysPeech: A Preprocessed Raw Audio Dataset For Speech Dysarthria is a Speech Dysarthria Dataset for the applicaton of Audio Classification, Speech Detection and similar avenues of research in ASR.
RAWDysPeech consists of raw audio files segregated into two classes: 1 and 0, where 1 is for speech involving Dysarthria and 0 is for normal speech. We combine and preprocess some of the most popular speech datasets available open sourced. TORGO, UASPEECH, Ultrax, EasyCall are a few to be named.
Here's a brief description of the steps taken to preprocess and combine and we also encourage you to cite the original authors if this dataset helps in your research.
--------------------------------------------------------
This dataset provides preprocessed speech recordings from the UASPEECH database, specifically enhanced for machine learning applications using advanced noise reduction and signal processing techniques.

Dataset Description
The dataset contains audio recordings that have been processed using:

I. FFT-based noise reduction:
Hanning window application for better frequency analysis
16-bit audio depth processing
44.1 kHz sampling rate[1]
Stereo channel support with dual MEMS microphone configuration[2]

II. Preprocessing Steps
Signal Processing

Background noise subtraction using ambient noise sampling
Frequency spectrum analysis with FFT
Amplitude scaling and normalization
Single-sided FFT amplitude doubling for accurate frequency representation[1]

III. Audio Parameters

Bit Depth: 16-bit (pyaudio.paInt16)
Sample Rate: 44.1 kHz
Buffer Size: 44100 frames
Channel Configuration: Supports both mono and stereo recording[2]

IV. File Format
Audio files are saved in .WAV format
Timestamps are included in filenames (YYYY_MM_DD_HH_MM_SS_pyaudio)
Data is organized in dedicated data folders with automated directory creation[1]

V. Applications
Speech Recognition Systems
Dysarthric Speech Analysis
Audio Classification Tasks
Speech Pattern Recognition
Acoustic Model Training
Technical Implementation
The preprocessing pipeline includes real-time audio capture, noise profiling, FFT analysis, and spectrogram generation, making it suitable for both research and practical applications
---------------------------------------------------
Citations:
[1] Heejin Kim, Mark Hasegawa Johnson, Jonathan Gunderson, Adrienne Perlman, Thomas Huang, Kenneth Watkin, Simone Frame, Harsh Vardhan Sharma, Xi Zhou, March 17, 2023, "UASpeech", IEEE Dataport, doi: https://dx.doi.org/10.21227/f9tc-ab45.
[2] Rudzicz, F., Namasivayam, A.K., Wolff, T. (2012) The TORGO database of acoustic and articulatory speech from speakers with dysarthria. Language Resources and Evaluation, 46(4), pages 523--541.
[3] Shah, Arya; Qureshi, Aymen; Polprasert, Chantri (2024), “ADAPTIVE: A Novel Dataset For Acoustic DysArthria deTection through temPoral Inference and Voice Engineering”, Mendeley Data, V1, doi: 10.17632/j5bgddf6rp.1

RAWDysPeech：一款面向构音障碍（Speech Dysarthria）的预处理原始音频数据集，是专为音频分类、语音检测以及自动语音识别（Automatic Speech Recognition，ASR）领域相关研究方向打造的构音障碍数据集。
RAWDysPeech包含按两类划分的原始音频文件：1和0，其中1代表构音障碍语音，0代表正常语音。本数据集整合并预处理了多款当前主流的开源语音数据集，例如TORGO、UASPEECH、Ultrax、EasyCall等。
下文将简要介绍数据集的预处理与整合流程，若本数据集对您的研究有所助益，敬请引用原始文献的作者。
--------------------------------------------------------
本数据集源自UASPEECH数据库的预处理语音录音，通过先进的降噪与信号处理技术进行增强优化，以适配机器学习应用场景。

数据集说明
本数据集包含经过以下处理的音频录音：
一、基于快速傅里叶变换（Fast Fourier Transform，FFT）的降噪处理：
采用汉宁窗（Hanning window）以优化频率分析效果；
16位音频深度处理；
44.1 kHz采样率[1]；
支持双MEMS麦克风配置的立体声通道[2]。

二、预处理步骤
信号处理环节：
通过环境噪声采样实现背景噪声消除；
基于FFT的频谱分析；
幅度缩放与归一化处理；
对单边FFT幅度进行加倍操作，以实现精准的频率表征[1]。

三、音频参数
比特深度：16位（对应pyaudio.paInt16）；
采样率：44.1 kHz；
缓冲区大小：44100帧；
通道配置：支持单声道与立体声录制[2]。

四、文件格式
音频文件以.WAV格式存储；
文件名中包含格式为YYYY_MM_DD_HH_MM_SS_pyaudio的时间戳；
数据通过自动创建的专用数据文件夹进行组织管理[1]。

五、应用场景
语音识别系统构建；
构音障碍语音分析；
音频分类任务；
语音模式识别；
声学模型训练。

技术实现
该预处理流水线涵盖实时音频采集、噪声轮廓分析、FFT分析以及语谱图生成，可同时适配科研与实际应用场景。
---------------------------------------------------
参考文献：
[1] Heejin Kim、Mark Hasegawa Johnson、Jonathan Gunderson、Adrienne Perlman、Thomas Huang、Kenneth Watkin、Simone Frame、Harsh Vardhan Sharma、Xi Zhou，2023年3月17日，《UASpeech》，IEEE数据仓库，DOI：https://dx.doi.org/10.21227/f9tc-ab45。
[2] Rudzicz, F.、Namasivayam, A.K.、Wolff, T. (2012) 构音障碍患者的声学与发音语音TORGO数据库. 《语言资源与评价》, 46(4), 第523--541页。
[3] Shah, Arya、Qureshi, Aymen、Polprasert, Chantri (2024), “ADAPTIVE: 一种基于时序推理与语音工程的声学构音障碍检测新型数据集”, Mendeley数据集, V1, DOI: 10.17632/j5bgddf6rp.1