Probably worth adjusting the title -- air quality is not just a measure of pollution. In this case, both SF and Portland are filled (unfortunately) with wildfire smoke, rather than a man-made pollutant...
Are you folks planning on extending this to speech? I'm always been disappointed by how speech vocoder networks aren't built with any great inductive biases for waveform generation (besides very long receptive fields), and have desperately wanted something like this tuned for speech. It'd be great if a DSP-based architecture could be shown to outperform WaveNet / Parallel WaveNet / WaveRNN / WaveFlow / etc, and I'd love to use that in our own work. (There's been some attempts based on source-filter models like the "neural source filter (NSF) network", but nothing's caught on as best as I can tell.)
I don't think anything about the current set of tools is specific to sample rate; WaveNet, Tacotron, WaveRNN, etc, should work fine to generate 44.1Khz audio. They might just need slightly different hyperparameters or sizes to work well, or may take longer to train due to longer sequence lengths.
There's a few recent papers actually that show minor improvements by integrating LPC prediction into deep methods ([0], [1]). In my experience (some of which comes from reproducing these, some of which comes from my own experiments), this isn't actually too useful, at at most offers a minor modeling benefit.
The main difference between something like Festival and what we have now is the amount of domain-specific engineering. (This is generally the promise of deep learning -- replace hand-engineered features with simple-to-understand features and a deep model.) If you go and read the Festival manual, you're going to find tons of domain-specific rules and heuristics and subroutines; for example, there's a page on writing letter to sound rules as a grammar [2]. Nowadays, we may have a pipeline that resembles Festival at the high level, but each step of the pipeline is learned as a deep model from data rather than being carefully hand-engineered by many people over the course of years. This yields much more fluid speech as well as much, much faster iteration and experimentation times, leading to faster progress as well.
First of all, it's important to note that Tacotron and WaveNet are responsible for different parts of the speech synthesis pipeline, so the comparison here isn't quite accurate. Specifically, Tacotron takes a representation of the text (characters, phonemes, etc) and converts it into a frame-level acoustic representation (spectrograms, log mel spectrograms, etc, spaced every 5-25ms). WaveNet takes a frame-level representation of the audio (for example, the output of Tacotron, or phonemes with frame-level timing information) and converts it to a waveform.
Second, I don't see any reason why there shouldn't be an open-source Tacotron or WaveNet implementation that's as good as Google's model implementations. Implementing and training these models is expensive but not prohibitively so (nowadays, you could probably do it with $5,000 - $10,000, including experimentation costs).
That said, quality of text-to-speech systems is determined only partially by the quality of these models -- much if not most of the work of building high quality text to speech systems goes into things like high quality data collection systems, good data annotations, good normalization and NLP tailored towards the domain of the TTS system, multilanguage support, optimized inference implementations for server or mobile platforms, etc.
This one is a bit tricky to answer since it requires defining "AI". AI as a moniker has been used to describe deep neural networks, search algorithms, expert systems and logic systems, particle filters, SVMs, etc etc. Almost all text-to-speech (TTS) systems are based on a combination of some of these machine-learning methods and digital signal processing (DPS), so I would say yeah, text-to-speech is exactly what AI describes, even if it doesn't resemble human-like thinking like other AI applications do.
2. Is there any active work in non-AI methods?
This one again is a bit tricky for the same reason as before. However, there's a ton of pieces of the TTS pipeline that aren't AI in the current sense of the word (machine learning with neural networks or HMMs or other classifiers). For example, concatenative systems will traditionally take a large database of audio, divide it into chunks, and then recombine those chunks, using some interpolation method such as (OLA, PSOLA) to overlap those chunks. Choosing the chunks to overlap to create the target speech becomes an AI / search problem, using some sort of acoustic model to predict the acoustic parameters of each frame and then using a Viterbi search algorithm with target / join costs to find the optimal chunks. As another example of non-AI parts of the pipeline, text normalization tends to involve a lot of hand-written rules; for example, should you say "5/10/2019" as "May tenth, twenty nineteen", "the tenth of may twenty nineteen", "the tenth of may two thousand nineteen", or even "october fifth twenty nineteen". This decision and the conversion is often done with a ton of handwritten rules or grammars (see Kestrel, Google's text normalization system, and the open-source version, cleverly named Sparrowhawk). Anyways, the real answer is that TTS is always a combination of AI (machine learning) approaches with specialized text and audio processing algorithms.
3. Which bits are the AI bits?
The AI bits are the bits where you need to make some sort of heuristic decisions, and you'd like to make them by imitating some target speech. For example, things like part of speech detection, predicting acoustic parameters (spectrograms, F0, etc), more recently waveform synthesis as well.
4. Do deep methods significantly improve on the state of the art?
Yes, though they also come at a cost. For example, deep sequence-to-sequence networks make great frame-level models: Tacotron and similar models can do things like emotional and stylized voice synthesis much better than what I've seen HMMs and other non-deep models do. As another example, WaveNet / WaveRNN / etc are some of the only parametric speech models (that is, generating the waveform from scratch instead of copying it from a database of audio) that can match the quality of concatenative models (copying audio from a database), but they can be quite difficult to deploy due to high computational cost. Overall, though, yeah, deep methods and all the improvements to neural networks in the past few years are having a profound impact on the quality and naturalness of TTS.
I'm an author on a few of these papers referenced (the Deep Voice papers from Baidu). I'm happy to answer any questions folks may have about neural speech synthesis, as I've been working on this for several years now.
In general, it's a fascinating space. There are challenges in text processing (not even mentioned in the blog), such as grapheme to phoneme conversion, part of speech detection, word sense disambiguation, text normalization, challenges in utterance-level modeling (spectrograms), and challenges in "spectrogram inversion" / waveform synthesis. The NLP components of the pipeline are often overlooked but are no less important than they were a few years ago -- part of speech / word sense is the difference between "Time is a CONstruct" and "I'm going to conSTRUCT a tower", and is the difference between "Let's drop that bass" being about a DJ or about a fish. The acoustic modeling phase (e.g. Tacotron, Deep Voice 3) works fairly well, and can produce some awesome demos with things like style tokens ("GST-Tacotron"), but still has a ways to go until it can encompass the full range of human inflection and emotion. At the waveform synthesis level, models like WaveRNN (with subscale modeling) and Parallel WaveNet make it possible to deploy modern waveform synthesis models, but it's still a major issue to deploy them onto low-power devices due to compute restrictions. Overall, lots of interesting challenges to work on, and we're making a lot of progress quite quickly; and I haven't even started talking about voice conversion or voice cloning!
The new Gmail interface has snooze, so if that's the only thing you're missing it'll stick around. I mostly miss the "Trips" feature, having recently switched from Inbox to new Gmail...
"Monoid" is an adjective that describes a data type.
Anything you describe as a monoid has to have three properties: you can add them together, there's an "empty" or "zero" value, and (a + b) + c = a + (b + c).
For example, strings with concatenation are a monoid, because you can use an empty string and the string + operator.
For example, integers with addition are a monoid, because you can use 0 and the integer + operator.
For example, sets are a monoid, because you can use the empty set and the union operator.
For example, a "Picture" data type could be a monoid, because you could have a fully-transparent picture and use an "overlay" operator to put one picture on top of another.
Why do you care? Well, for one thing, it just makes it easy to find the function or operator you want to use: as the OP said, if I feel like my data type should have a concat or append operator but I don't know what it's called, I just use mappend / mempty / mconcat.
But, as another practical example: "sum" in Python works for numbers, but I've seen folks assume that it works for anything that you can use + on, and so try and use strings with it. But, because "sum" isn't a general purpose tool, that doesn't work! In Haskell, on the other hand, the equivalent -- mconcat -- will work on anything that is a monoid. (And can be specialized to work faster for specific data structures.)
Other languages also have monoids. If someone tells you Haskell has monoids and other languages don't, what they really mean is, Haskell makes explicit the monoid pattern / interface, where other languages have it implicitly for different data types. Talking about the pattern explicitly isn't revolutionary, but it can be pretty useful for discoverability and writing "general" algorithms.
Instead of calling data structures monoids, you'd get the same effect if, as a programming language community, you decided that as many types as possible should support "+", and that every data type that can should have a makeEmptyThing() method, and that it would be weird and not ok if x + makeEmptyThing() didn't equal x and if (x + y) + z didn't equal x + (y + z), and then subtly shamed any libraries that defined makeEmptyThing() and + in ways that didn't follow that pattern. But if your programming language community did this (for consistency) you'd probably want to come up with a name for it -- "Addable", "Concatable", etc -- and the Haskell community chose "Monoid" (because of relationships to theory etc etc).
To be fair, this article is really pretty "cutting edge" as far as Haskell goes. It describing a new feature they're planning on adding to the Haskell compiler, one that significantly extends the type system in an interesting way. I think it's fair for them to assume comfort and familiarity with Haskell syntax and code when writing this article; their audience is folks who are likely to know about and use somewhat esoteric GHC extensions...
Feel free to get in touch for more Q/A, my email is in my profile.
We've experimented a bunch with many of these hyperparameters. Our phoneme signal has mostly stayed 256 Hz, but we've done a few experiments with lower-frequency signals that indicate it's probably possible to reduce it.
We have used many types of upsampling, and find that the upsampling and conditioning procedure does not affect the quality of the audio itself, but does affect the frequency of pronunciation mistakes. We used bicubic and bilinear interpolation based upsampling, as well as transposed convolutions and a variety of other simpler convolutions (for example, per-channel transposed convolutions). These tend to work and converge, but then generate pronunciation mistakes on difficult phonemes. A full transposed convolution upsampling (two transposed convolution layers with stride 8 each) works almost as well as our bidirectional QRNNs, but it's much, much, more expensive in terms of compute and parameters, and takes longer to train as well.
As noted in the paper, we used many of the original features used for WaveNet before reducing our feature set. F0 is definitely important for proper intonation. We find that including the surrounding phonemes is quite important; with the bidirectional QRNN upsampling, leaving those out still works, but not nearly as well. It seems likely that a different conditioning network would remove the need for those "context" phonemes.
We have not yet used an encoder-decoder approach for duration or F0. Char2Wav has a bunch of interesting ideas, and it may be a direction for our future work. However, we do not plan on including the grapheme-to-phoneme model into our main model, because it's crucial that we easily affect the pronunciation of phonemes with a phoneme dictionary; by having an explicit grapheme-to-phoneme step, we can easily set the pronunciation for unseen words (like "P!nk" or "Worcestershire"; an integrated grapheme-to-phoneme model would not be able to do those, even humans usually cannot!).
We have not yet worked with speaker global conditioning, but it is likely that the results from the WaveNet paper apply to our WaveNet implementation as well.
Finally, as for sampling, we have not seen much variation due to random seed for a fully converged model. However, our intuition for why sampling is important is that the speech distribution is (a) multimodal and (b) biased towards silence. If you are interested, you can gain a little bit of intuition about what the distribution actually looks like by just plotting a color map across time, with high-probability values being bright and low probability values being dark; it generates a pretty plot, and you can see that some areas are clearly stochastic (especially fricatives) and some areas are multimodal (vowel wave peaks).
I don't quite know what VoCo does, but it seems like a concatenative system that they've tuned a huge amount. I'm a little skeptical that it works as well and as reliably in real life as it does in demos. But, even so, there parametric models tend to be much smaller in size and more flexible, so there may be applications where WaveNet-style systems are applicable in ways concatenative systems can't handle (high quality on-device TTS, emotive TTS, speaker synthesis for new unheard speakers, etc).
Check out Char2Wav (recent) and SampleRNN (the RNN-based audio synthesis architecture). The related work section of the Deep Voice paper mention a bunch of related papers that are relevant!
We take several days (2-3) on 8 Titan X GPUs to train our models, which is quite a lot of compute. Running on mobile devices is quite challenging – the inference is not yet fast enough to support that, and has only been optimized for x86 AVX2 CPUs. It may be possible with a fair amount of future work!
It's hard to say! We don't quite know exactly how many parameters or minutes of audio are needed to describe fully someone's voice and speaking patterns. Maybe one or two, maybe much more.
Thank you for clarifying this! We tried fairly hard to make this clear, because as you say, the hard part is generating inflection and duration that sounds natural. There's still a ton of work left to do in this duration – we're clearly nowhere near being able to generate human-level speech.
Our work is meant to make working with TTS easier to deep learning researchers by describing a complete and trainable system that can be trained completely from data, and demonstrate that the neural vocoder substitutes can actually be deployed to streaming production servers. Future work (both by us and hopefully other groups) will make further progress for inflection synthesis!
Right now, we do not have plans to make an API available. This paper and blog post are mostly meant to describe our techniques to other deep learning researchers and spur innovation in the field. However, we hope that these techniques will be available eventually, and we'll provide more information when that happens.
Baidu as a company doesn't use TensorFlow (as far as I know). We have our own high-performance and easy to use open source framework called PaddlePaddle [0], which is quite powerful and flexible.
However, the Baidu Research Silicon Valley AI Lab (SVAIL) allows researchers and research teams to use whatever frameworks they want to, and we have projects using TensorFlow, Torch, our own SVAIL-internal internal high-performance RNN framework, and PaddlePaddle. Using our own framework sometimes allows us to work on very high-performance implementations of various primitives and techniques that would be harder to do without complete control over the source code.