Machine Learning Speech Recognition

Keeping up my yearly blogging cadence, it’s about time I wrote to let people know what I’ve been up to for the last year or so at Mozilla. People keeping up would have heard of the sad news regarding the Connected Devices team here. While I’m sad for my colleagues and quite disappointed in how this transition period has been handled as a whole, thankfully this hasn’t adversely affected the Vaani project. We recently moved to the Emerging Technologies team and have refocused on the technical side of things, a side that I think most would agree is far more interesting, and also far more suited to Mozilla and our core competence.

Project DeepSpeech

So, out with Project Vaani, and in with Project DeepSpeech (name will likely change…) – Project DeepSpeech is a machine learning speech-to-text engine based on the Baidu Deep Speech research paper. We use a particular layer configuration and initial parameters to train a neural network to translate from processed audio data to English text. You can see roughly how we’re progressing with that here. We’re aiming for a 10% Word Error Rate (WER) on English speech at the moment.

You may ask, why bother? Google and others provide state-of-the-art speech-to-text in multiple languages, and in many cases you can use it for free. There are multiple problems with existing solutions, however. First and foremost, most are not open-source/free software (at least none that could rival the error rate of Google). Secondly, you cannot use these solutions offline. Third, you cannot use these solutions for free in a commercial product. The reason a viable free software alternative hasn’t arisen is mostly down to the cost and restrictions around training data. This makes the project a great fit for Mozilla as not only can we use some of our resources to overcome those costs, but we can also use the power of our community and our expertise in open source to provide access to training data that can be used openly. We’re tackling this issue from multiple sides, some of which you should start hearing about Real Soon Now™.

The whole team has made contributions to the main code. In particular, I’ve been concentrating on exporting our models and writing clients so that the trained model can be used in a generic fashion. This lets us test and demo the project more easily, and also provides a lower barrier for entry for people that want to try out the project and perhaps make contributions. One of the great advantages of using TensorFlow is how relatively easy it makes it to both understand and change the make-up of the network. On the other hand, one of the great disadvantages of TensorFlow is that it’s an absolute beast to build and integrates very poorly with other open-source software projects. I’ve been trying to overcome this by writing straight-forward documentation, and hopefully in the future we’ll be able to distribute binaries and trained models for multiple platforms.

Getting Involved

We’re still at a fairly early stage at the moment, which means there are many ways to get involved if you feel so inclined. The first thing to do, in any case, is to just check out the project and get it working. There are instructions provided in READMEs to get it going, and fairly extensive instructions on the TensorFlow site on installing TensorFlow. It can take a while to install all the dependencies correctly, but at least you only have to do it once! Once you have it installed, there are a number of scripts for training different models. You’ll need a powerful GPU(s) with CUDA support (think GTX 1080 or Titan X), a lot of disk space and a lot of time to train with the larger datasets. You can, however, limit the number of samples, or use the single-sample dataset (LDC93S1) to test simple code changes or behaviour.

One of the fairly intractable problems about machine learning speech recognition (and machine learning in general) is that you need lots of CPU/GPU time to do training. This becomes a problem when there are so many initial variables to tweak that can have dramatic effects on the outcome. If you have the resources, this is an area that you can very easily help with. What kind of results do you get when you tweak dropout slightly? Or layer sizes? Or distributions? What about when you add or remove layers? We have fairly powerful hardware at our disposal, and we still don’t have conclusive results about the affects of many of the initial variables. Any testing is appreciated! The Deep Speech 2 paper is a great place to start for ideas if you’re already experienced in this field. Note that we already have a work-in-progress branch implementing some of these ideas.

Let’s say you don’t have those resources (and very few do), what else can you do? Well, you can still test changes on the LDC93S1 dataset, which consists of a single sample. You won’t be able to effectively tweak initial parameters (as unsurprisingly, a dataset of a single sample does not represent the behaviour of a dataset with many thousands of samples), but you will be able to test optimisations. For example, we’re experimenting with model quantisation, which will likely be one of multiple optimisations necessary to make trained models usable on mobile platforms. It doesn’t particularly matter how effective the model is, as long as it produces consistent results before and after quantisation. Any optimisation that can be made to reduce the size or the processor requirement of training and using the model is very valuable. Even small optimisations can save lots of time when you start talking about days worth of training.

Our clients are also in a fairly early state, and this is another place where contribution doesn’t require expensive hardware. We have two clients at the moment. One written in Python that takes advantage of TensorFlow serving, and a second that uses TensorFlow’s native C++ API. This second client is the beginnings of what we hope to be able to run on embedded hardware, but it’s very early days right now.

And Finally

Imagine a future where state-of-the-art speech-to-text is available, for free (in cost and liberty), on even low-powered devices. It’s already looking like speech is going to be the next frontier of human-computer interaction, and currently it’s a space completely tied up by entities like Google, Amazon, Microsoft and IBM. Putting this power into everyone’s hands could be hugely transformative, and it’s great to be working towards this goal, even in a relatively modest capacity. This is the vision, and I look forward to helping make it a reality.

Open Source Speech Recognition

I’m currently working on the Vaani project at Mozilla, and part of my work on that allows me to do some exploration around the topic of speech recognition and speech assistants. After looking at some of the commercial offerings available, I thought that if we were going to do some kind of add-on API, we’d be best off aping the Amazon Alexa skills JS API. Amazon Echo appears to be doing quite well and people have written a number of skills with their API. There isn’t really any alternative right now, but I actually happen to think their API is quite well thought out and concise, and maps well to the sort of data structures you need to do reliable speech recognition.

So skipping forward a bit, I decided to prototype with Node.js and some existing open source projects to implement an offline version of the Alexa skills JS API. Today it’s gotten to the point where it’s actually usable (for certain values of usable) and I’ve just spent the last 5 minutes asking it to tell me Knock-Knock jokes, so rather than waste any more time on that, I thought I’d write this about it instead. If you want to try it out, check out this repository and run npm install in the usual way. You’ll need pocketsphinx installed for that to succeed (install sphinxbase and pocketsphinx from github), and you’ll need espeak installed and some skills for it to do anything interesting, so check out the Alexa sample skills and sym-link the ‘samples‘ directory as a directory called ‘skills‘ in your ferris checkout directory. After that, just run the included example file with node and talk to it via your default recording device (hint: say ‘launch wise guy‘).

Hopefully someone else finds this useful – I’ll be using this as a base to prototype further voice experiments, and I’ll likely be extending the Alexa API further in non-standard ways. What was quite neat about all this was just how easy it all was. The Alexa API is extremely well documented, Node.js is also extremely well documented and just as easy to use, and there are tons of libraries (of varying quality…) to do what you need to do. The only real stumbling block was pocketsphinx’s lack of documentation (there’s no documentation at all for the Node bindings and the C API documentation is pretty sparse, to say the least), but thankfully other members of my team are much more familiar with this codebase than I am and I could lean on them for support.

I’m reasonably impressed with the state of lightweight open source voice recognition. This is easily good enough to be useful if you can limit the scope of what you need to recognise, and I find the Alexa API is a great way of doing that. I’d be interested to know how close the internal implementation is to how I’ve gone about it if anyone has that insider knowledge.

Web Navigation Transitions

Wow, so it’s been over a year since I last blogged. Lots has happened in that time, but I suppose that’s a subject for another post. I’d like to write a bit about something I’ve been working on for the last week or so. You may have seen Google’s proposal for navigation transitions, and if not, I suggest reading the spec and watching the demonstration. This is something that I’ve thought about for a while previously, but never put into words. After reading Google’s proposal, I fear that it’s quite complex both to implement and to author, so this pushed me both to document my idea, and to implement a proof-of-concept.

I think Google’s proposal is based on Android’s Activity Transitions, and due to Android UI’s very different display model, I don’t think this maps well to the web. Just my opinion though, and I’d be interested in hearing peoples’ thoughts. What follows is my alternative proposal. If you like, you can just jump straight to a demo, or view the source. Note that the demo currently only works in Gecko-based browsers – this is mostly because I suck, but also because other browsers have slightly inscrutable behaviour when it comes to adding stylesheets to a document. This is likely fixable, patches are most welcome.


 Navigation Transitions specification proposal

Abstract

An API will be suggested that will allow transitions to be performed between page navigations, requiring only CSS. It is intended for the API to be flexible enough to allow for animations on different pages to be performed in synchronisation, and for particular transition state to be selected on without it being necessary to interject with JavaScript.

Proposed API

Navigation transitions will be specified within a specialised stylesheet. These stylesheets will be included in the document as new link rel types. Transitions can be specified for entering and exiting the document. When the document is ready to transition, these stylesheets will be applied for the specified duration, after which they will stop applying.

Example syntax:

When navigating to a new page, the current page’s ‘transition-exit‘ stylesheet will be referenced, and the new page’s ‘transition-enter‘ stylesheet will be referenced.

When navigation is operating in a backwards direction, by the user pressing the back button in browser chrome, or when initiated from JavaScript via manipulation of the location or history objects, animations will be run in reverse. That is, the current page’s ‘transition-enter‘ stylesheet will be referenced, and animations will run in reverse, and the old page’s ‘transition-exit‘ stylesheet will be referenced, and those animations also run in reverse.

[Update]

Anne van Kesteren suggests that forcing this to be a separate stylesheet and putting the duration information in the tag is not desirable, and that it would be nicer to expose this as a media query, with the duration information available in an @-rule. Something like this:

I think this would indeed be nicer, though I think the exact naming might need some work.

Transitioning

When a navigation is initiated, the old page will stay at its current position and the new page will be overlaid over the old page, but hidden. Once the new page has finished loading it will be unhidden, the old page’s ‘transition-exit‘ stylesheet will be applied and the new page’s ‘transition-enter’ stylesheet will be applied, for the specified durations of each stylesheet.

When navigating backwards, the CSS animations timeline will be reversed. This will have the effect of modifying the meaning of animation-direction like so:

and this will also alter the start time of the animation, depending on the declared total duration of the transition. For example, if a navigation stylesheet is declared to last 0.5s and an animation has a duration of 0.25s, when navigating backwards, that animation will effectively have an animation-delay of 0.25s and run in reverse. Similarly, if it already had an animation-delay of 0.1s, the animation-delay going backwards would become 0.15s, to reflect the time when the animation would have ended.

Layer ordering will also be reversed when navigating backwards, that is, the page being navigated from will appear on top of the page being navigated backwards to.

Signals

When a transition starts, a ‘navigation-transition-startNavigationTransitionEvent will be fired on the destination page. When this event is fired, the document will have had the applicable stylesheet applied and it will be visible, but will not yet have been painted on the screen since the stylesheet was applied. When the navigation transition duration is met, a ‘navigation-transition-end‘ will be fired on the destination page. These signals can be used, amongst other things, to tidy up state and to initialise state. They can also be used to modify the DOM before the transition begins, allowing for customising the transition based on request data.

JavaScript execution could potentially cause a navigation transition to run indefinitely, it is left to the user agent’s general purpose JavaScript hang detection to mitigate this circumstance.

Considerations and limitations

Navigation transitions will not be applied if the new page does not finish loading within 1.5 seconds of its first paint. This can be mitigated by pre-loading documents, or by the use of service workers.

Stylesheet application duration will be timed from the first render after the stylesheets are applied. This should either synchronise exactly with CSS animation/transition timing, or it should be longer, but it should never be shorter.

Authors should be aware that using transitions will temporarily increase the memory footprint of their application during transitions. This can be mitigated by clear separation of UI and data, and/or by using JavaScript to manipulate the document and state when navigating to avoid keeping unused resources alive.

Navigation transitions will only be applied if both the navigating document has an exit transition and the target document has an enter transition. Similarly, when navigating backwards, the navigating document must have an enter transition and the target document must have an exit transition. Both documents must be on the same origin, or transitions will not apply. The exception to these rules is the first document load of the navigator. In this case, the enter transition will apply if all prior considerations are met.

Default transitions

It is possible for the user agent to specify default transitions, so that navigation within a particular origin will always include navigation transitions unless they are explicitly disabled by that origin. This can be done by specifying navigation transition stylesheets with no href attribute, or that have an empty href attribute.

Note that specifying default transitions in all situations may not be desirable due to the differing loading characteristics of pages on the web at large.

It is suggested that default transition stylesheets may be specified by extending the iframe element with custom ‘default-transition-enter‘ and ‘default-transition-exit‘ attributes.

Examples

Simple slide between two pages:

[page-1.html]

[page-1-exit.css]

[page-2.html]

[page-2-enter.css]


I believe that this proposal is easier to understand and use for simpler transitions than Google’s, however it becomes harder to express animations where one element is transitioning to a new position/size in a new page, and it’s also impossible to interleave contents between the two pages (as the pages will always draw separately, in the predefined order). I don’t believe this last limitation is a big issue, however, and I don’t think the cognitive load required to craft such a transition is considerably higher. In fact, you can see it demonstrated by visiting this link in a Gecko-based browser (recommended viewing in responsive design mode Ctrl+Shift+m).

I would love to hear peoples’ thoughts on this. Am I actually just totally wrong, and Google’s proposal is superior? Are there huge limitations in this proposal that I’ve not considered? Are there security implications I’ve not considered? It’s highly likely that parts of all of these are true and I’d love to hear why. You can view the source for the examples in your browser’s developer tools, but if you’d like a way to check it out more easily and suggest changes, you can also view the git source repository.

Efficient animation for games on the (mobile) web

Drawing on some of my limited HTML5 games experience, and marginally less limited general games and app writing experience, I’d like to write a bit about efficient animation for games on the web. I usually prefer to write about my experiences, rather than just straight advice-giving, so I apologise profusely for how condescending this will likely sound. I’ll try to improve in the future 🙂

There are a few things worth knowing that will really help your game (or indeed app) run better and use less battery life, especially on low-end devices. I think it’s worth getting some of these things down, as there’s evidence to suggest (in popular and widely-used UI libraries, for example) that it isn’t necessarily common knowledge. I’d also love to know if I’m just being delightfully/frustratingly naive in my assumptions.

First off, let’s get the basic stuff out of the way.

Help the browser help you

If you’re using DOM for your UI, which I’d certainly recommend, you really ought to use CSS transitions and/or animations, rather than JavaScript-powered animations. Though JS animations can be easier to express at times, unless you have a great need to synchronise UI animation state with game animation state, you’re unlikely to be able to do a better job than the browser. The reason for this is that CSS transitions/animations are much higher level than JavaScript, and express a very specific intent. Because of this, the browser can make some assumptions that it can’t easily make when you’re manually tweaking values in JavaScript. To take a concrete example, if you start a CSS transition to move something from off-screen so that it’s fully visible on-screen, the browser knows that the related content will end up completely visible to the user and can pre-render that content. When you animate position with JavaScript, the browser can’t easily make that same assumption, and so you might end up causing it to draw only the newly-exposed region of content, which may introduce slow-down. There are signals at the beginning and end of animations that allow you to attach JS callbacks and form a rudimentary form of synchronisation (though there are no guarantees on how promptly these callbacks will happen).

Speaking of assumptions the browser can make, you want to avoid causing it to have to relayout during animations. In this vein, it’s worth trying to stick to animating only transform and opacity properties. Though some browsers make some effort for other properties to be fast, these are pretty much the only ones semi-guaranteed to be fast across all browsers. Something to be careful of is that overflow may end up causing relayouting, or other expensive calculations. If you’re setting a transform on something that would overlap its container’s bounds, you may want to set overflow: hidden on that container for the duration of the animation.

Use requestAnimationFrame

When you’re animating canvas content, or when your DOM animations absolutely must synchronise with canvas content animations, do make sure to use requestAnimationFrame. Assuming you’re running in an arbitrary browsing session, you can never really know how long the browser will take to draw a particular frame. requestAnimationFrame causes the browser to redraw and call your function before that frame gets to the screen. The downside of using this vs. setTimeout, is that your animations must be time-based instead of frame-based. i.e. you must keep track of time and set your animation properties based on elapsed time. requestAnimationFrame includes a time-stamp in its callback function prototype, which you most definitely should use (as opposed to using the Date object), as this will be the time the frame began rendering, and ought to make your animations look more fluid. You may have a callback that ends up looking something like this:

You’ll note that I set startTime to -1 at the beginning, when I could just as easily set the time using the Date object and avoid the extra code in the animation callback. I do this so that any setup or processes that happen between the start of the animation and the callback being processed don’t affect the start of the animation, and so that all the animations I start before the frame is processed are synchronised.

To save battery life, it’s best to only draw when there are things going on, so that would mean calling requestAnimationFrame (or your refresh function, which in turn calls that) in response to events happening in your game. Unfortunately, this makes it very easy to end up drawing things multiple times per frame. I would recommend keeping track of when requestAnimationFrame has been called and only having a single handler for it. As far as I know, there aren’t solid guarantees of what order things will be called in with requestAnimationFrame (though in my experience, it’s in the order in which they were requested), so this also helps cut out any ambiguity. An easy way to do this is to declare your own refresh function that sets a flag when it calls requestAnimationFrame. When the callback is executed, you can unset that flag so that calls to that function will request a new frame again, like this:

Following this pattern, or something similar, means that no matter how many times you call requestRedraw, your drawing function will only be called once per frame.

Remember, that when you do drawing in requestAnimationFrame (and in general), you may be blocking the browser from updating other things. Try to keep unnecessary work outside of your animation functions. For example, it may make sense for animation setup to happen in a timeout callback rather than a requestAnimationFrame callback, and likewise if you have a computationally heavy thing that will happen at the end of an animation. Though I think it’s certainly overkill for simple games, you may want to consider using Worker threads. It’s worth trying to batch similar operations, and to schedule them at a time when screen updates are unlikely to occur, or when such updates are of a more subtle nature. Modern console games, for example, tend to prioritise framerate during player movement and combat, but may prioritise image quality or physics detail when compromise to framerate and input response would be less noticeable.

Measure performance

One of the reasons I bring this topic up, is that there exist some popular animation-related libraries, or popular UI toolkits with animation functions, that still do things like using setTimeout to drive their animations, drive all their animations completely individually, or other similar things that aren’t conducive to maintaining a high frame-rate. One of the goals for my game Puzzowl is for it to be a solid 60fps on reasonable hardware (for the record, it’s almost there on Galaxy Nexus-class hardware) and playable on low-end (almost there on a Geeksphone Keon). I’d have liked to use as much third party software as possible, but most of what I tried was either too complicated for simple use-cases, or had performance issues on mobile.

How I came to this conclusion is more important than the conclusion itself, however. To begin with, my priority was to write the code quickly to iterate on gameplay (and I’d certainly recommend doing this). I assumed that my own, naive code was making the game slower than I’d like. To an extent, this was true, I found plenty to optimise in my own code, but it go to the point where I knew what I was doing ought to perform quite well, and I still wasn’t quite there. At this point, I turned to the Firefox JavaScript profiler, and this told me almost exactly what low-hanging-fruit was left to address to improve performance. As it turned out, I suffered from some of the things I’ve mentioned in this post; my animation code had some corner cases where they could cause redraws to happen several times per frame, some of my animations caused Firefox to need to redraw everything (they were fine in other browsers, as it happens – that particular issue is now fixed), and some of the third party code I was using was poorly optimised.

A take-away

To help combat poor animation performance, I wrote Animator.js. It’s a simple animation library, and I’d like to think it’s efficient and easy to use. It’s heavily influenced by various parts of Clutter, but I’ve tried to avoid scope-creep. It does one thing, and it does it well (or adequately, at least). Animator.js is a fire-and-forget style animation library, designed to be used with games, or other situations where you need many, synchronised, custom animations. It includes a handful of built-in tweening functions, the facility to add your own, and helper functions for animating object properties. I use it to drive all the drawing updates and transitions in Puzzowl, by overriding its requestAnimationFrame function with a custom version that makes the request, but appends the game’s drawing function onto the end of the callback, like so:

My game’s redraw function does all drawing, and my animation callbacks just update state. When I request a redraw outside of animations, I just check the animator’s activeAnimations property first to stop from mistakenly drawing multiple times in a single animation frame. This gives me nice, synchronised animations at very low cost. Puzzowl isn’t out yet, but there’s a little screencast of it running on a Nexus 5:

Alternative, low-framerate YouTube link.