CONVERSATIONAL SPEECH DATASETTEXT → STYLE → SPEECH

PromptDialog

with Open-Ended
Text Descriptions

A large-scale conversational speech dataset with open-ended style descriptions. Bringing the nuance of natural language to multi-speaker speech synthesis.

BEYOND PREDEFINED TAGS
TEXT

“This is not my day.”

Open-ended Description

Choke out the words with a voice breaking with despair.

8,503hConversational speech
79%Utterances with style descriptions
Open-ended StyleStyle beyond a fixed set of tags

01 / THE IDEA

Aligning style descriptions, dialogue text, and audiobook speech

Comparison of dataset construction pipelines: our pipeline aligns source narrative with audiobook speech and extracts open-ended descriptions; the conventional pipeline converts annotated attribute tags into descriptions.
FIGURE 01 ↗

Comparison of pipelines for constructing speech datasets with natural-language descriptions. In (1), the open-ended examples are drawn from our dataset, and the bold text highlights open-ended style elements that are difficult for external tools to annotate.

02 / PROMPTDIALOG DATASET

Dataset samples

Listen to Chinese and English samples from PromptDialog alongside the corresponding dialogue and narrative descriptions.

Loading dataset samples…