WaveNet: A Generative Model for Raw Audio · HackerPapers discussions

WaveNet: A Generative Model for Raw Audio — mentioned by rhdunn

Fri, 16 Jan 2026 10:21:13 GMT

Using neural nets (machine learning) to train TTS voices has been around a long time. [1] (2016 https://arxiv.org/abs/1609.03499) WaveNet: A Generative Model for Raw Audio [2] (2017 https://arxiv.org/abs/1711.10433) Parallel WaveNet: Fast High-Fidelity Speech Synthesis [3] (2021 https://arxiv.org/abs/2106.07889) UnivNet: A Neural Vocoder with Multi-Resolution Spectrogram Discriminators for High-Fidelity Waveform Generation [4] (2022 https://arxiv.org/abs/2203.14941) Neural Vocoder is All You Need for Speech Super-resolution

Paper and discussion history. A mention is not an endorsement.

WaveNet: A Generative Model for Raw Audio — mentioned by rryan

Mon, 22 Aug 2022 04:14:46 GMT

You're using a pointwise loss, which means you are treating each timepoint as conditionally independent. That's a deeply flawed assumption, probabilistically. Try an autoregressive model of the joint probability distribution of the sine wave timepoints -- like WaveNet [1]. It will nail your sine wave -- just as it nails mixtures of sine waves (speech and music). :) [1] https://arxiv.org/abs/1609.03499

Paper and discussion history. A mention is not an endorsement.

WaveNet: A Generative Model for Raw Audio — mentioned by zip6

Thu, 18 Apr 2019 22:59:49 GMT

I don't know what the current state of the art is, but you should take a look at WaveNet if you haven't already - https://arxiv.org/pdf/1609.03499.pdf While trying to search for an answer, I came across this list of papers, which has a section on speech synthesis - https://github.com/zzw922cn/awesome-speech-recognition-speec...

Paper and discussion history. A mention is not an endorsement.

WaveNet: A Generative Model for Raw Audio — mentioned by probably_wrong

Tue, 30 Oct 2018 18:25:56 GMT

The simplest way would be downloading something like MaryTTS [1], read the documentation, and train your own voice model. It won't be perfect, but shouldn't be too hard. The best results would probably be achieved by implementing DeepMind's WaveNet paper [2], but it might be too much for what you need. I'm not really sure what to suggest in between those two. Some kind of convolutional NN, I guess? [1] mary.dfki.de [2] arxiv.org/abs/1609.03499

Paper and discussion history. A mention is not an endorsement.

WaveNet: A Generative Model for Raw Audio — mentioned by tree_of_item

Wed, 15 Aug 2018 16:17:50 GMT

Isn't this like an "auto-regressive feed forward neural net"[0]? > Instead of making predictions from a state that depends on the entire history, an autoregressive model directly predicts yt using only the k most recent inputs, xt−k+1,…,xt. This corresponds to a strong conditional independence assumption. In particular, a feed-forward model assumes the target only depends on the k most recent inputs. Google’s WaveNet[1] nicely illustrates this general principle. [0] http://bair.berkeley.edu/blog/2018/08/06/recurrent/ [1] https://arxiv.org/abs/1609.03499

Paper and discussion history. A mention is not an endorsement.

WaveNet: A Generative Model for Raw Audio — mentioned by jdietrich

Sun, 20 May 2018 12:45:15 GMT

Duplex's underlying text-to-speech technology research (WaveNet) has produced several papers and is now in public beta. It represents a huge advance in text-to-speech fidelity, using a remarkably straightforward algorithm. https://arxiv.org/pdf/1609.03499.pdf https://www.isca-speech.org/archive/Interspeech_2017/pdfs/14... https://arxiv.org/pdf/1712.05884.pdf https://cloud.google.com/text-to-speech/

Paper and discussion history. A mention is not an endorsement.

WaveNet: A Generative Model for Raw Audio — mentioned by ml_thoughts

Sun, 08 Oct 2017 00:05:10 GMT

I think an aspect of deep learning that is often overlooked is that it is still not clear how much of current algorithm performance is defined by local "obsession to detail" vs global "understanding" of the subject matter. The "tiramisu" layers here are an interesting example of this: they are built on dilated convolutions and one of their main selling points is that they can do calculations on a pixel-by-pixel basis, as opposed to standard methods which use are forced to compress information along the way through pooling/strided convolutions (basically taking multiple pixels and summarizing…

Paper and discussion history. A mention is not an endorsement.

WaveNet: A Generative Model for Raw Audio — mentioned by option

Tue, 09 May 2017 23:17:38 GMT

Yes - https://arxiv.org/pdf/1609.03499.pdf

Paper and discussion history. A mention is not an endorsement.

WaveNet: A Generative Model for Raw Audio — mentioned by kastnerkyle

Wed, 08 Mar 2017 17:06:46 GMT

Showing one of the top two samples from their blog (full prediction) along with the one you link (only acoustic model) would more clearly show what you are getting at in the explanation, since how the text inference works is most of the complexity in this model (given baseline knowledge of WaveNet, at least). The sample shown only does step 3 from your summary as far as I can tell. In particular the top 2 samples from the Baidu blog most clearly show the gaps we need to cover from a research perspective to truly get "human level" TTS - a lot of the complexity in TTS is in the text part, and…

Paper and discussion history. A mention is not an endorsement.

WaveNet: A Generative Model for Raw Audio — mentioned by bmc7505

Sat, 26 Nov 2016 01:09:57 GMT

> Did the original WaveNet text to speech demo come with a paper or source code? Paper, yes. [1] Source code, no. [1] https://arxiv.org/pdf/1609.03499.pdf

Paper and discussion history. A mention is not an endorsement.