CFM是一种新技术,已被证明可以改进扩散模型,Meta的Voicebox模型将CFM引入语音合成领域,下面是voicebox的一个工作流程图
Matcha-TTS是第一个开源conditional normalising flows语音合成项目,提供基于 LJSpeech 和 VCTK 数据预训练模型以供测评
Matcha-TTS有两个主要的贡献和其他建议:
1. We propose an improved encoder-decoder TTS architecture that uses a combination of 1D CNNs and Transformers in the decoder. This reduces memory consumption and is fast to evaluate, improving synthesis speed.
相对于Grad-TTS的decoder,使用了1D CNNs替换2D CNNs、并加入Transformers块
2. We train these models using optimal-transport conditional flow matching (OT-CFM) , which is a new method to learn ODEs that sample from a data distribution. Compared to conventional CNFs and score-matching probability flow ODEs, OT-CFM defines simpler paths from source to target, enabling accurate synthesis in fewer steps than DPMs.
使用Flow Matching加速技术
3. 使用旋转位置编码(rotational position embeddings)?RoPE,减少存储
4. 使用MAS对齐
5. 使用snake beta激活函数
??开源地址:
https://github.com/shivammehta25/Matcha-TTS
??工程展示:
https://shivammehta25.github.io/Matcha-TTS/
??在线推理:
https://huggingface.co/spaces/shivammehta25/Matcha-TTS
??中文实现:
https://github.com/PlayVoice/Grad-TTS-Chinese?
(Grad-TTS-CFM,其他优化还未集成)
模型架构:
性能指标:
推理界面:
中文测试句子:
时光仿佛有穿越到了从前,在你诗情画意的眼波中,在你舒适浪漫的暇思里,我如风中的思绪徜徉广阔天际,仿佛一片沾染了快乐的羽毛,在云环影绕颤动里浸润着风的呼吸,风的诗韵,那清新的耳语,那婉约的甜蜜,那恬淡的温馨,将一腔情澜染得愈发的缠绵。(Grad-TTS-CFM,使用BigVGAN通用声码器,优化1&3&5还未集成,还有明显发音错误)