DreamVoiceDB

Timbre annotations and natural-language voice descriptions for 900 speakers, released with the DreamVoice text-guided voice conversion paper (Interspeech 2024).

📄 Paper (arXiv:2406.16314) · 💻 Code · 🔊 Demo

Overview

Most voice conversion systems need a reference recording of the target speaker. DreamVoice instead lets a user describe the voice they want in plain text, for example "a warm, deep voice of an older man, well suited to storytelling." Training models like this needs speech paired with reliable descriptions of how each voice sounds, and no such resource existed.

DreamVoiceDB fills that gap. It provides:

• Expert timbre annotations for 900 speakers from LibriTTS-R and VCTK. The annotations cover age, gender, brightness, roughness, perceived qualities (e.g., warm, strong, authoritative) and suitability for professional speaking roles.
• Natural-language voice prompts generated from those annotations. There are multiple descriptions per speaker, covering different combinations of their attributes.

To our knowledge, DreamVoiceDB was the first publicly released dataset of text-based timbre annotations for text-guided voice conversion.

Note: This repository contains annotations only. The audio must be obtained separately from LibriTTS-R and VCTK under their respective licenses.

Dataset structure

The repository has two CSV files with different schemas, exposed as two configurations.

tags: one row per speaker (DreamVoiceDB_speaker_and_tags.csv)

Column	Values	Description
speaker_id	e.g. 1348, p240	Speaker ID in the source corpus. Numeric IDs are LibriTTS-R; p/s-prefixed IDs are VCTK.
gender	Male, Female, Ambiguous	Perceived gender of the voice
age	Teenager, Adult, Senior	Perceived age group
dark_bright	Dark, Bright, blank	Perceived brightness
rough_smooth	Rough, Smooth, blank	Perceived roughness
attractive	Attractive, blank	Subjective quality
authoritative	Authoritative, blank	Subjective quality
cute	Cute, blank	Subjective quality
nasal	Nasal, blank	Subjective quality
strong	Strong, blank	Subjective quality
warm	Warm, blank	Subjective quality
weak	Weak, blank	Subjective quality
profession	list of: Storytelling, Public Presentation, Client and Public Interaction, Diplomacy and Judiciary	Speaking roles the voice is judged well suited for. Stored as a Python-style list string; may be empty ([]).

A blank attribute cell means the attribute was not assigned to that speaker, i.e., annotators did not reach sufficient agreement that it applies. It does not mean the opposite quality.

Example:

speaker_id,gender,age,dark_bright,rough_smooth,attractive,authoritative,cute,nasal,strong,warm,weak,profession
1638,Male,Senior,Dark,,,Authoritative,,,Strong,Warm,,['Storytelling']

prompts: multiple rows per speaker (DreamVoiceDB_speaker_and_prompts.csv)

Column	Description
speaker_id	Speaker ID; joins to the tags config
combination	The subset of the speaker's attributes used to write this prompt (e.g., Female, Senior, Storytelling)
if_finished	Prompt-generation status flag (1 = generated)
prompt	Natural-language description of the voice

Example:

126,"Female, Senior, Diplomacy and Judiciary",1,"Her aged, feminine voice is well-suited for political negotiation and legal discourse in the field of diplomacy and judiciary."

Usage

from datasets import load_dataset

tags = load_dataset("carankt/DreamVoiceDB", "tags", split="train")
prompts = load_dataset("carankt/DreamVoiceDB", "prompts", split="train")

print(tags[0])
print(prompts[0])

Or with pandas:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download("carankt/DreamVoiceDB", "DreamVoiceDB_speaker_and_prompts.csv", repo_type="dataset")
prompts = pd.read_csv(path)

To pair prompts with audio, map speaker_id to the matching speaker folder in LibriTTS-R or VCTK.

Tip: when evaluating generalization to unseen voices, split by speaker_id rather than by prompt, so the same speaker never appears in both training and test sets.

How the dataset was built

Annotation followed a three-stage process designed with input from an expert voice actor:

1. Keyword selection. A vocabulary of timbre keywords was selected and split into two groups. The objective group is age, gender, brightness and roughness. The subjective group covers perceived qualities such as strength, warmth and authority, plus suitability for professional speaking roles.
2. Expert survey. Eight trained annotators labeled every speaker. Their backgrounds were in speech-language pathology, speech and accent coaching, singing coaching, and transcription, and each was selected through a pilot test run. Brightness and roughness were rated on expert-defined Likert scales against reference audio; subjective qualities were rated on a binary scale.
3. Agreement filtering. Keywords with unanimous agreement were kept. Keywords with moderate agreement were reassessed using the distribution of annotator votes and self-reported confidence scores.

Prompt generation. For each speaker, natural-language descriptions were generated with GPT-4 from combinations of the speaker's verified keywords, including leave-one-out and leave-many-out subsets. This gives each voice many descriptions with varied wording and level of detail.

Models trained on DreamVoiceDB

The DreamVoice paper uses this dataset to train two systems:

• DreamVC: an end-to-end, diffusion-based voice conversion model conditioned on the text prompt.
• DreamVG: a lightweight plugin that generates a speaker embedding from text, and works with existing one-shot voice conversion models (e.g., FreeVC).

Code and checkpoints are in the DreamVoice repository.

Limitations

• Labels are perceptual and subjective. They describe how listeners heard each voice, not the speaker's actual identity, age or gender.
• Speakers come from read-speech English corpora (audiobooks and studio recordings), so accents, ages and speaking styles are limited to those sources.
• Prompts are LLM-generated from the keyword labels and may contain stylistic repetition.
• The source audio is not redistributed here; check the LibriTTS-R and VCTK licenses before using or sharing it.

Citation

If you use DreamVoiceDB, please cite:

@inproceedings{hai2024dreamvoice,
  title     = {DreamVoice: Text-Guided Voice Conversion},
  author    = {Hai, Jiarui and Thakkar, Karan and Wang, Helin and Qin, Zengyi and Elhilali, Mounya},
  booktitle = {Proc. Interspeech 2024},
  year      = {2024}
}

Jiarui Hai and Karan Thakkar contributed equally to this work.

Authors and contact

Jiarui Hai*, Karan Thakkar*, Helin Wang, Zengyi Qin, Mounya Elhilali Johns Hopkins University · Massachusetts Institute of Technology · MyShell.ai *Equal contribution

Questions: open a discussion on this repository, or contact kthakka2@jhu.edu.