File size: 1,996 Bytes
929ed1b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
This document outlines the protocols and procedures used during the data collection phase of the Speech Recognition Bias Reduction Project.

## 1. Research Overview
The study aims to gather voice recordings from non-native English speakers to enhance automated speech recognition systems' accuracy. The dataset will include audio recordings of participants spelling names, reading phone numbers, and reciting addresses, along with relevant metadata.

## 2. Participant Recruitment
- **Target Population:** International and Native students at Duke University with diverse linguistic backgrounds.
- **Recruitment Channels:** Participants were recruited through email lists, student groups, and social media platforms.

## 3. Informed Consent Process
Participants were provided with a detailed consent form explaining the study's purpose, data usage, and their rights as participants. Consent was obtained through a simple text input ("Yes") within the survey.

## 4. Recording Process
- **Platform:** Audio recordings were collected via the Qualtrics Survey App, accessible on both laptops and mobile devices.
- **Recording Instructions:** Participants were given clear instructions on how to record their audio:
  - Spell out a randomized name letter by letter.
  - Read out a randomized phone number digit by digit.
  - Recite a randomized address clearly.

## 5. Audio Quality and Environment
Participants were instructed to record their audio in their natural environments to capture realistic conditions, which may include background noise. The recordings aimed to reflect typical conditions under which users interact with automated voice systems.

## 6. Data Management
- **Anonymization:** All participant data, including audio recordings and metadata, is anonymized to protect participant identities before public release.
- **Withdrawal Option:** Participants were informed that they could withdraw their recordings before the dataset is made publicly accessible.