Text-to-Speech2017en

The LJ Speech Dataset

13,100 short audio clips of a single speaker reading passages from non-fiction books. Standard benchmark for single-speaker TTS.

Metrics:mos, mcd

Paper / Website Download

Current State of the Art

VALL-E 2

Microsoft

4.61

mos

LJ Speech — mos

5 results · 3 SOTA advances · higher is better

All results

SOTA frontier

mos Progress Over Time

Showing 3 breakthroughs from Jun 2021 to Jun 2024

Key Milestones

Jun 2021

VITS

MOS (1–5). VITS end-to-end TTS. Source: Table 2, arxiv:2106.06103 (ICML 2021)

4.4

Source

Jun 2023

StyleTTS 2

MOS (1–5). Surpasses human baseline (4.44 MOS). Source: Table 2, arxiv:2306.07279 (NeurIPS 2023)

4.5

+2.7%

Source

Jun 2024

VALL-E 2Current SOTA

MOS (1–5). Human parity: CMOS +0.17 above ground truth. Source: Table 1, arxiv:2406.05370 (Jun 2024)

4.6

+1.3%

Source

Total Improvement

4.1%

Time Span

3y 1m

Breakthroughs

Current SOTA

4.6

Top Models Performance Comparison

Top 5 models ranked by mos

Best Score

4.6

Top Model

VALL-E 2

Models Compared

Score Range

0.400

mosPrimary

#	Model	Score	Paper / Code	Date
1	VALL-E 2 Microsoft	4.61	VALL-E 2: Neural Codec Language Models are Human Parity Zero-Shot Text to Speech Synthesizers	Jun 2024
2	StyleTTS 2Open Source Columbia University	4.55	StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models	Jun 2023
3	VITSOpen Source Kakao	4.43	VITS: Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech	Jun 2021
4	Voicebox Meta AI	4.3	Voicebox: Text-Guided Multilingual Universal Speech Generation at Scale	Jun 2023
5	XTTS v2Open Source Coqui AI	4.21	XTTS: A Massively Multilingual Zero-Shot Text-to-Speech Model	Apr 2023