“It is still surprising to me how good the audio models from WaveNet and DeepVoice are (such as the sample you show)”
WaveNet: A Generative Model for Raw Audio
WaveNet:逐采样点自回归生成原始音频波形的深度网络
WaveNet 是一个直接生成原始音频波形的深度神经网络,完全概率化且自回归,每个采样点的预测分布取决于此前所有采样点,并且能在每秒数万采样点的数据上高效训练。用于语音合成时,英语和普通话的自然度均被人类听众评为显著优于当时最好的参数式和拼接式系统;单个模型可通过说话人条件切换不同说话人,训练于音乐时也能生成逼真的片段,还可用作判别模型做音素识别。
评论者把它当作神经语音合成的里程碑,称赞其保真度和算法的简洁,同时也指出它一次只能看到约 300 毫秒的上下文,难以学到长程结构。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“It represents a huge advance in text-to-speech fidelity, using a remarkably straightforward algorithm.”
“you should take a look at WaveNet if you haven't already”
“The compromise is Wavenet, which can only act on a context of about ~300 ms at a time [2], which more or less precludes learning long-range structure”
“Try an autoregressive model of the joint probability distribution of the sine wave timepoints -- like WaveNet [1].”
“Using neural nets (machine learning) to train TTS voices has been around a long time.”