File size: 2,468 Bytes
90c07f0
abd9a73
90c07f0
 
 
 
 
 
60c520b
6a53d44
 
 
 
 
 
 
abd9a73
 
90c07f0
 
d8dc492
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
---
title: TTSDS Benchmark and Leaderboard
emoji: 🥇
colorFrom: green
colorTo: indigo
sdk: gradio
app_file: app.py
pinned: true
license: mit
tags:
  - leaderboard
  - submission:semiautomatic
  - test:public
  - judge:auto
  - modality:audio
  - eval:generation
  - tts
short_description: Text-To-Speech (TTS) Evaluation using objective metrics.
---

# TTSDS Benchmark

As many recent Text-to-Speech (TTS) models have shown, synthetic audio can be close to real human speech.
However, traditional evaluation methods for TTS systems need an update to keep pace with these new developments.
Our TTSDS benchmark assesses the quality of synthetic speech by considering factors like prosody, speaker identity, and intelligibility.
By comparing these factors with both real speech and noise datasets, we can better understand how synthetic speech stacks up.

## More information
More details can be found in our paper [*TTSDS -- Text-to-Speech Distribution Score*](https://arxiv.org/abs/2407.12707).

## Reproducibility
To reproduce our results, check out our repository [here](https://github.com/ttsds/ttsds).

## Credits


This benchmark is inspired by [TTS Arena](https://huggingface.co/spaces/TTS-AGI/TTS-Arena) which instead focuses on the subjective evaluation of TTS models.
Our benchmark would not be possible without the many open-source TTS models on Hugging Face and GitHub.
Additionally, our benchmark uses the following datasets:
- [LJSpeech](https://keithito.com/LJ-Speech-Dataset/h)
- [LibriTTS](https://www.openslr.org/60/)
- [VCTK](https://datashare.ed.ac.uk/handle/10283/2950)
- [Common Voice](https://commonvoice.mozilla.org/)
- [ESC-50](https://github.com/karolpiczak/ESC-50)
And the following metrics/representations/tools:
- [Wav2Vec2](https://arxiv.org/abs/2006.11477)
- [Hubert](https://arxiv.org/abs/2006.11477)
- [WavLM](https://arxiv.org/abs/2110.13900)
- [PESQ](https://en.wikipedia.org/wiki/Perceptual_Evaluation_of_Speech_Quality)
- [VoiceFixer](https://arxiv.org/abs/2204.05841)
- [WADA SNR](https://www.cs.cmu.edu/~robust/Papers/KimSternIS08.pdf)
- [Whisper](https://arxiv.org/abs/2212.04356)
- [Masked Prosody Model](https://huggingface.co/cdminix/masked_prosody_model)
- [PyWorld](https://github.com/JeremyCCHsu/Python-Wrapper-for-World-Vocoder)
- [WeSpeaker](https://arxiv.org/abs/2210.17016)
- [D-Vector](https://github.com/yistLin/dvector)

Authors: Christoph Minixhofer, Ondřej Klejch, and Peter Bell 
of the University of Edinburgh.