This post on Deep Voice seems a little off-the-mark. In fact, I would say it is completely misleading about the technical accomplishments here.
From my perspective, Baidu's approach is a little embarrassing, with the use of many modeling stages in their training and production of TTS. When the rest of the community is moving towards end-to-end training, their usage of this many stages sounds excruciating. Merlin[0], which was a pretty good standard for 2016, has this painful feeling as well, with two DL stages (duration, acoustic) followed by some conditioning and then a synthesis step.
The more important technical contribution seems to be the hand-tuned synthesis code that makes their generation faster; cool but not particularly sexy (and there are few details). The details on training hyperparams are nice to have too, of course.
Contrary to the post, I would be very surprised if the voice sample included in the post was actually generated by Deep Voice -- it has none of the robotic qualities pointed out by the researchers themselves in their blog post[1]. More likely it is a demonstration of the loss in their last, WaveNet-like stage. This was also pointed out in the previous HN discussion[2]
Lastly, Andrew Ng is neither thanked in the paper nor mentioned on any webpage -- are we sure this was work he supervised?
- I state the caveats that the voice sample published for Deep Voice are using ground truth features. That being said I can make it clearer.
- Andrew Ng runs the Baidu AI team. He may not have supervised it but he's associated with this.
- I've gotten direct feedback from an original author of this paper to ensure the post represents its accomplishments well. At this point I believe it does save for the caveat.
Are there any how-to guides to getting something like this running? Or are there missing pieces of closed-source software needed? I'm interested in the theory but it helps to be able to try it out.
If you want to see the most basic form of this pipeline, I have a blog post on "bad speech synthesis" [0]. There are open source versions of WaveNet for TTS [1], but I have not run the code myself or seen the quality of the output. Our code for char2wav is theoretically available [2], but not yet ready for "how-to-guide" level use.
Microsoft's speech libraries are easily accessible thru a .NET-clr language, such as C# or Powershell. I made a voice-bot as a weekend project last year: https://news.ycombinator.com/item?id=11663029
Dhruv's planning on putting out a series of posts, where he'll probably be including the model training/set up components! Look out for the next bits this coming week.
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 getting the subtleties of stress and f0 prediction correct is far from a solved problem. This is partly why there are so many different submodules and parts in the text piece of DeepVoice, and a whole appendix dedicated to that in WaveNet.
It is still surprising to me how good the audio models from WaveNet and DeepVoice are (such as the sample you show) - it seems that given good enough text features e.g. groundtruth the synthesis is nearly perfect. So it seems (IMO) that the next research papers will be focused on the text/f0 part.
I will also plug the work we have been doing at MILA, which is related but tries to directly go from text -> speech with attention based RNNs [0]. A longer version of our paper should be on arxiv soon, but for now we have a short teaser which was submitted as an ICLR workshop [1]. One fun feature is that we have the ability to handle multi-speaker synthesis, and multi-speaker datasets as well.
The primary differences from a high level are: we don't need a pronunciation dictionary for training, but they (DeepVoice or WaveNet) do. However we need a vocoder (currently WORLD) to get intermediate targets for training, while they do not. For training we need a directory of text files, and a directory of audio files, along with a script to run WORLD on the audio.
This means char2wav can go to new languages without extra lexical information (e.g. pronunciation dictionaries), since the WORLD assumptions are broad, but we find we have a difficult time on English without using phonemes (from a pronunciation dict), since the character -> audio mapping is pretty difficult for English.
We have better outputs (compared to char2wav English) on Spanish, German, and Romanian [2] though we are still improving every day. Note that the Romanian is still using fixed synthesis with WORLD - this is a demo only of the text -> WORLD features part of the model, the "reader".
The end result of both char2wav and DeepVoice (be able to go from text -> speech directly in generation) is the same. WaveNet also has this - you can see most of the TTS details in the appendix of the paper [3].
There is also a good discussion on reddit with one of the DeepVoice authors that has quite a bit of detail about their approach, I found it quite helpful [4].
Heiga Zen has a great overview talk on youtube for people who are interested in TTS [5].
There are also some interesting extensions on WaveNet [6][7], though I have not translated enough to determine what exact methods are being used.
From my perspective, Baidu's approach is a little embarrassing, with the use of many modeling stages in their training and production of TTS. When the rest of the community is moving towards end-to-end training, their usage of this many stages sounds excruciating. Merlin[0], which was a pretty good standard for 2016, has this painful feeling as well, with two DL stages (duration, acoustic) followed by some conditioning and then a synthesis step.
The more important technical contribution seems to be the hand-tuned synthesis code that makes their generation faster; cool but not particularly sexy (and there are few details). The details on training hyperparams are nice to have too, of course.
Contrary to the post, I would be very surprised if the voice sample included in the post was actually generated by Deep Voice -- it has none of the robotic qualities pointed out by the researchers themselves in their blog post[1]. More likely it is a demonstration of the loss in their last, WaveNet-like stage. This was also pointed out in the previous HN discussion[2]
Lastly, Andrew Ng is neither thanked in the paper nor mentioned on any webpage -- are we sure this was work he supervised?
[0] https://github.com/CSTR-Edinburgh/merlin
[1] http://research.baidu.com/deep-voice-production-quality-text...
[2] https://news.ycombinator.com/item?id=13756489