DreamVoice Dataset

Introduction

The DreamVoice dataset is designed to support research in text-guided voice conversion and synthetic voice generation. The dataset is structured into three main files:

dreamvoice_s1_raw_label.csv: Contains raw keyword votes from multiple experts. Each row represents the collective opinions of experts regarding various aspects of the speaker's voice characteristics.
dreamvoice_s2_keyword.csv: Provides verified keywords for each speaker, serving as a reliable reference for their voice profile based on expert consensus.
dreamvoice_s3_synthetic_prompt.csv: Includes synthetic timbre description prompts. These prompts are designed to describe the unique timbre characteristics of each speaker, facilitating the generation of synthetic voices that match specific timbral qualities.

Usage

To use the DreamVoice dataset in your research or projects, follow these steps:

Download VCTK and LibriTTS-R Datasets
The DreamVoice dataset is built to work with the VCTK and LibriTTS-R speech datasets. You can download them from the following links:
- VCTK Dataset
- LibriTTS-R Dataset
Working with the Dataset
Once the datasets are downloaded, you can combine them with the DreamVoice dataset based on the column speaker_id.

DreamVoice Dataset Details

Survey Method

Figure 1: Schematic diagram of DreamVoiceDB survey method.

Annotation Strategy

Datasets Used: LibriTTS-R and VCTK
Number of Speakers: 900 [100 from VCTK and 800 from LibriTTS-R, sampled randomly]
Keyword Selection Process:
- Guided by an expert voice actor
- Total of 10 keywords identified
- Keywords categorized based on relative objectivity and subjectivity
Annotator Profile:
- Total Annotators: 8 (4 females and 4 males)
- Expertise: Speech pathology, accent coaching, singing coaching, transcription, and translation

You can find our sample survey page used for data collection here - DreamVoiceDB Survey Link

Keywords and Descriptors

The DreamVoiceDB dataset utilizes a variety of keywords and descriptors to annotate voice timbre. These are grouped into relatively objective and subjective categories, along with additional aspects related to suitability for various voice-related professions.

Relatively Objective Keywords [Multtple Options Single Choice]

Age: 'Teenager or Young', 'Adult or Neutral', and 'Senior or Old'.
Gender: 'Definitely Female', 'Somewhat Ambiguous / Neutral'.
Brightness: Ranging from 'Very Dark (Rich)' to 'Very Bright (Thin)'.
Roughness: From 'Rough (Raspy)' to 'Smooth (Mellow)'.

Reference Audio can be found in the following powerpoint presentation: DreamVoiceDB Reference Audio

Relatively Subjective Keywords [Multiple Options Multiple Choice]

Nasal: A noticeable resonance or vibration in the nasal passages.
Breathy: You can hear the flow of air as the person speaks.
Weak: Lack of energy and confidence.
Strong: Full, resonant, and confident.
Cute: Perceived as young and endearing.
Attractive: A voice that draws pleasure in listeners.
Cool: Laid-back, confident, and collected.
Warm: A voice that conveys kindness, sincerity, and comfort.
Authoritative: A voice that carries weight and command.
Others: Other characteristics but not related to the speaker's emotion.

Additional Aspects

This includes evaluating the voice's fit for professions such as singing, acting, public speaking, and other vocations where vocal qualities are paramount. Specific examples include:

Public Presentation: Public Speaker, Motivational Speaker, Teacher, Tour Guide
Storytelling: Podcaster, Audiobook Narrator
Client and Public Interaction: Customer Service, Healthcare Communicator, Retail Associate, Receptionist
Diplomacy and Judiciary: Diplomat, Judge
None of the above

Each keyword and descriptor plays a crucial role in accurately capturing and conveying the unique characteristics of each voice sample in the dataset.

Analysis and Enhancement

Agreement scores used to prioritize keywords
Rigorous reassessment for moderately agreed keywords
For each speaker we atleast have age and gender keywords
Other keywords are selected based on the Agreement scores
We do not use any fine grained annotations for darkness and roughness in the model training
Use of OpenAI's GPT4 for natural language descriptors
Generation of descriptors based on all keyword combinations

Generated Descriptors Examples:

Speaker ID	WAV File	Keywords	Prompt
2562	File 2562	Male, Adult, Dark	An adult male voice, with a dark timbre.
30	File 30	Female, Teenager, Bright	A teenage girl's voice, radiating brightness and energy.
2156	File 2156	Male, Senior, Strong	A senior male voice, characterized by strength and power.

Dataset Statistics

Figure 2: Dataset statistics.

image imagewidth (px) 2.07k 4.8k