TTTS Arena-AI 声音模型板
与20+文本对语音模型相比,官方基准、社区评级和并肩比较。
逐边比较
模型板
| # | 型 型 | 正式官方官方官方官方官方 | 社区 社区 | 你的评分 | 速度速度 | 级别 |
|---|---|---|---|---|---|---|
| 1 |
Kokoro
Lightweight 82M parameter model delivering studio-quality speech with blazing-fast inference.
82M
1200h
2024
|
4.8 /5 |
5.0
/5
1 vote
|
fast | Free | |
| 2 |
CosyVoice 2
Alibaba's scalable streaming TTS with human-parity naturalness and near-zero latency.
300M
200000h
2024
|
4.26 /5 | 尚未表决 | medium | Standard | |
| 3 |
Chatterbox
State-of-the-art zero-shot voice cloning with emotion control from Resemble AI.
300M
2025
|
4.25 /5 | 尚未表决 | medium | Premium | |
| 4 |
StyleTTS 2
Human-level text-to-speech through style diffusion and adversarial training.
100M
585h
2024
|
4.23 /5 | 尚未表决 | medium | Premium | |
| 5 |
Piper
A fast, local neural text to speech system optimized for Raspberry Pi and embedded devices.
15M
2023
|
4.15 /5 | 尚未表决 | fast | Free | |
| 6 |
MeloTTS
High-quality multilingual text-to-speech that runs on CPU with minimal latency.
25M
2024
|
4.13 /5 | 尚未表决 | fast | Free | |
| 7 |
Dia TTS
Multi-speaker dialog generation model that creates natural conversations between speakers.
1.6B
2024
|
4.09 /5 | 尚未表决 | medium | Standard | |
| 8 |
VITS
Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech.
25M
585h
2021
|
4.0 /5 | 尚未表决 | fast | Free | |
| 9 |
Orpheus
Human-level emotional TTS model trained on 100K hours of speech data.
3B
100000h
2025
|
4.0 /5 | 尚未表决 | medium | Standard | |
| 10 |
OpenVoice
Instant voice cloning with granular control over style, emotion, and accent.
300M
2024
|
4.0 /5 | 尚未表决 | medium | Premium | |
| 11 |
IndexTTS-2
Zero-shot TTS with fine-grained emotion control and high expressiveness.
300M
2025
|
3.91 /5 | 尚未表决 | medium | Standard | |
| 12 |
Spark TTS
Voice cloning TTS with controllable emotion and speaking style via prompts.
500M
2025
|
3.9 /5 | 尚未表决 | medium | Standard | |
| 13 |
Parler TTS
Describe the voice you want in natural language and Parler generates matching speech.
880M
45000h
2024
|
3.83 /5 | 尚未表决 | medium | Standard | |
| 14 |
Tortoise TTS
Multi-voice text-to-speech focused on quality with autoregressive architecture.
400M
50000h
2022
|
3.7 /5 | 尚未表决 | slow | Premium | |
| 15 |
Bark
Transformer-based text-to-audio model that generates realistic speech, music, and sound effects.
350M
100000h
2023
|
3.57 /5 | 尚未表决 | slow | Standard | |
| 16 |
Bark Small
Lighter version of Bark with faster inference and lower memory usage.
150M
100000h
2023
|
— | 尚未表决 | medium | Standard | |
| 17 |
GLM-TTS
Achieves the lowest character error rate among open-source TTS models.
300M
2025
|
— | 尚未表决 | medium | Standard | |
| 18 |
GPT-SoVITS
Few-shot voice cloning TTS that replicates any voice from just 5 seconds of audio.
200M
2024
|
— | 尚未表决 | slow | Standard | |
| 19 |
Qwen3 TTS
Alibaba's multilingual TTS with voice cloning, preset voices, and voice design from text.
1.7B
2025
|
— | 尚未表决 | medium | Standard | |
| 20 |
Sesame CSM
Conversational speech model generating natural dialogue with appropriate timing and emotion.
1B
2025
|
— | 尚未表决 | slow | Premium | |
| 21 |
Chatterbox Turbo
Faster Chatterbox with sub-200ms latency and paralinguistic tags for laughs, coughs, and more.
350M
2025
|
— | 尚未表决 | fast | Standard | |
| 22 |
Dia 2
Streaming-first conversational TTS with multi-speaker dialogue and paralinguistic cues.
2B
2025
|
— | 尚未表决 | fast | Standard | |
| 23 |
VoxCPM
Tokenizer-free TTS producing 44.1kHz audio with context-aware paragraph consistency.
500M
1800000h
2025
|
— | 尚未表决 | fast | Standard | |
| 24 |
OuteTTS
LLM-based TTS that runs on CPU, GPU, or browser via llama.cpp and Transformers.js.
1B
5000h
2025
|
— | 尚未表决 | fast | Free | |
| 25 |
TADA
Zero-hallucination TTS with text-acoustic dual alignment, 5x faster than comparable LLM TTS.
1B
2026
|
— | 尚未表决 | fast | Standard | |
| 26 |
VibeVoice
Microsoft's multi-speaker long-form TTS generating up to 90 minutes with 4 distinct speakers.
1.5B
100000h
2025
|
— | 尚未表决 | fast | Standard | |
| 27 |
Pocket TTS
Lightweight 100M parameter model by Kyutai with voice cloning from a single sample.
100M
50000h
2025
|
— | 尚未表决 | fast | Free | |
| 28 |
Kitten TTS
Ultra-lightweight TTS under 80MB. Runs on CPU without GPU.
80M
2025
|
— | 尚未表决 | fast | Free | |
| 29 |
CosyVoice3
Next-generation multilingual TTS with bi-streaming, emotion control, and zero-shot voice cloning.
500M
200000h
2025
|
— | 尚未表决 | fast | Standard | |
| 30 |
MOSS-TTS
Ultra-long 20-language TTS supporting up to 1 hour of continuous generation with phoneme-level control.
8B
500000h
2026
|
— | 尚未表决 | medium | Premium | |
| 31 |
MegaTTS3
ByteDance's sparse alignment TTS with adjustable intelligibility vs. speaker similarity.
1B
100000h
2025
|
— | 尚未表决 | slow | Premium |
详细基准分数
Official TTS.ai benchmark scores across three dimensions: naturalness, accuracy, and speed.
Kokoro
Free
自然性质
4.8/5
准确性
4.7/5
速度速度
4.9/5
总体 总体
4.8/5
CosyVoice 2
Standard
自然性质
4.5/5
准确性
4.4/5
速度速度
3.8/5
总体 总体
4.26/5
Chatterbox
Premium
自然性质
4.7/5
准确性
4.5/5
速度速度
3.4/5
总体 总体
4.25/5
StyleTTS 2
Premium
自然性质
4.5/5
准确性
4.3/5
速度速度
3.8/5
总体 总体
4.23/5
Piper
Free
自然性质
3.5/5
准确性
4.2/5
速度速度
4.95/5
总体 总体
4.15/5
MeloTTS
Free
自然性质
3.8/5
准确性
4.1/5
速度速度
4.6/5
总体 总体
4.13/5
Dia TTS
Standard
自然性质
4.6/5
准确性
4.3/5
速度速度
3.2/5
总体 总体
4.09/5
VITS
Free
自然性质
3.4/5
准确性
4.0/5
速度速度
4.8/5
总体 总体
4.0/5
Orpheus
Standard
自然性质
4.3/5
准确性
4.1/5
速度速度
3.5/5
总体 总体
4.0/5
OpenVoice
Premium
自然性质
4.0/5
准确性
4.1/5
速度速度
3.9/5
总体 总体
4.0/5
IndexTTS-2
Standard
自然性质
4.3/5
准确性
4.1/5
速度速度
3.2/5
总体 总体
3.91/5
Spark TTS
Standard
自然性质
4.2/5
准确性
4.0/5
速度速度
3.4/5
总体 总体
3.9/5
Parler TTS
Standard
自然性质
4.1/5
准确性
3.9/5
速度速度
3.4/5
总体 总体
3.83/5
Tortoise TTS
Premium
自然性质
4.6/5
准确性
4.4/5
速度速度
1.8/5
总体 总体
3.7/5
Bark
Standard
自然性质
4.2/5
准确性
3.8/5
速度速度
2.5/5
总体 总体
3.57/5
基准基准方法
测试设置
- 硬件 : NVIDIA Tesla P40(各24GBVRAM),96GB总计
- 测试文本 : 5个标准化段落,涵盖不同的演讲模式(叙述、对话、技术、情感、多语言)
- 评价: 自动度量(MOS估计、WER、RTF)与人体听觉测试相结合
- 运行 : 每个模型每通过每次测试10次,平均分数d
标分标准
- 自然(40%): 欢乐、共鸣、节奏、情感——它听起来如何?
- 准确性(30%): 发音正确性、字出错率、可知性
- 速度(30%): 实时因数(音量秒/每一代秒)更高=更快。
- 总体情况: 加权平均数:0.4x自然+0.3x精确度+0.3x速度
注: 基准反映了我们具体硬件和测试文本的绩效。 现实世界的质量可能因输入文本、语言和语音选择而不同。 社区评级提供了基于不同实际用途的补充信号。
常问问题
我们能改进什么?您的反馈帮助我们解决问题。