🔈TTS技术部署与对比
主要对比内容:
- 生成质量:整体流畅即可。
- 生成速度:Real Time Factor(RTF)至少小于 0.5
- 推理要求:CPU 还是 GPU,VRAM 需求多少。
名称 | 离线 | 质量 | 速度 RTF | 推理要求 | 可控性 | 链接 |
---|---|---|---|---|---|---|
Edge TTS | ❌ | 0.7 | ||||
CosyVoice-python SDKs | ❌ |
|||||
CosyVoice-websocket | ❌ |
https://help.aliyun.com/zh/isi/developer-reference/websocket-protocol-description?spm=a2c4g.11186623.0.0.748e5757uTmjSr | ||||
Fish TTS | ⭕ | GPU 4GB | https://speech.fish.audio/ | |||
Ekho | ⭕ |
传统音素拼接 | <0.2 | CPU 7MB | / | https://www.eguidedog.net/ekho.php |
Melo TTS | ⭕ |
官方:CPU 实时推理 | https://github.com/myshell-ai/MeloTTS | |||
ChatTTS | ⭕ |
4090 0.3 | 占用有点大,GPU 推理。4GB | |||
tensorflowTTS-Tactron | ⭕ |
130MB | https://github.com/Z-yq/TensorflowTTS | |||
Paddle 工具包-fastspeech 2 | https://github.com/PaddlePaddle/PaddleSpeech?tab=readme-ov-file | |||||
GPT‐SoVITS‐v2 | 效果自然 | 大概率不会快 | ||||
parlertts-880M | 大概率也不快 | |||||
mars 5-tts 750+450M 参数 | ||||||
emotivoice | ||||||
SummerTTS | 效果还行,挺自然 | 1 | CPU | |||
Piper-Chinese | 韵律不太好 | 0.09 | 小 | |||
FCH-TTS (Parallel TTS) | 只有英文,要重新训练中文 ` | https://github.com/atomicoo/FCH-TTS |
||||