Low-latency audio on Windows with GStreamer

Digital audio is so ubiquitous that we rarely stop to think or wonder how the gears turn underneath our all-pervasive apps for entertainment. Today we'll look at one specific piece of the machinery: latency.

Let's say you're making a video of someone's birthday party with an app on your phone. Once the recording starts, you don't care when the app starts writing it to disk—as long as everything is there in the end.

However, if you're having a Skype call with your friend, it matters a whole lot how long it takes for the video to reach the other end and vice versa. It's impossible to have a conversation if the lag (latency) is too high.

The difference is, do you need real-time feedback or not?

Other examples, in order of increasingly stricter latency requirements are: live video streaming, security cameras, augmented reality games such as Pokémon Go, multiplayer video games in general, audio effects apps for live music recording, and many many more.

“But Nirbheek”, you might ask, “why doesn't everyone always ‘immediately’ send/store/show whatever is recorded? Why do people have to worry about latency?” and that's a great question!

To understand that, checkout my previous blog post, Latency in Digital Audio. It's also a good primer on analog vs digital audio!

Low latency on consumer operating systems

Each operating system has its own set of application APIs for audio, and each has a lower bind on the achievable latency:

• Linux has alsa-lib (old), Pulseaudio (standard), JACK (pro-audio), and Pipewire (under development)
• macOS and iOS have CoreAudio (standard, pro-audio)
• Android has AudioFlinger (Java API, android.media), OpenSL ES (C/C++ API), and AAudio (C/C++ API, new, pro-audio)
• Windows has DirectSound (deprecated), WASAPI (standard), and ASIO (proprietary, old, pro-audio).
• BSDs still use OSS

GStreamer already has plugins for almost all of these¹ (plus others that aren't listed here), and on Windows, GStreamer has been using the DirectSound API by default for audio capture and output since the very beginning.

However, the DirectSound API was deprecated in Windows XP, and with Vista, it was removed and replaced with an emulation layer on top of the newly-released WASAPI. As a result, the plugin can't be configured to have less than 200ms of latency, which makes it unsuitable for all the low-latency use-cases mentioned above. The DirectSound API is quite crufty and unnecessarily complex anyway.

GStreamer is rarely used in video games, but it is widely used for live streaming, audio/video calls, and other real-time applications. Worse, the WASAPI GStreamer plugins were effectively untouched and unused since the initial implementation in 2008 and were completely broken².

This left no way to achieve low-latency audio capture or playback on Windows using GStreamer.

The situation became particularly dire when GStreamer added a new implementation of the WebRTC spec in this release cycle. People that try it out on Windows were going to see much higher latencies than they should.

Luckily, I rewrote most of the WASAPI plugin code in January and February, and it should now work well on all versions of Windows from Vista to 10! You can get binary installers for GStreamer or build it from source.

Shared and Exclusive WASAPI

WASAPI allows applications to open sound devices in two modes: shared and exclusive. As the name suggests, shared mode allows multiple applications to output to (or capture from) an audio device at the same time, whereas exclusive mode does not.

Almost all applications should open audio devices in shared mode. It would be quite disastrous if your YouTube videos played without sound because Spotify decided to open your speakers in exclusive mode.

In shared mode, the audio engine has to resample and mix audio streams from all the applications that want to output to that device. This increases latency because it must maintain its own audio ringbuffer for doing all this, from which audio buffers will be periodically written out to the audio device.

In theory, hardware mixing could be used if the sound card supports it, but very few sound cards implement that now since it's so cheap to do in software. On Windows, only high-end audio interfaces used for professional audio implement this.

Another option is to allocate your audio engine buffers directly in the sound card's memory with DMA, but that complicates the implementation and relies on good drivers from hardware manufacturers. Microsoft has tried similar approaches in the past with DirectSound and been burned by it, so it's not a route they took with WASAPI³.

On the other hand, some applications know they will be the only ones using a device, and for them all this machinery is a hindrance. This is why exclusive mode exists. In this mode, if the audio driver is implemented correctly, the application's buffers will be directly written out to the sound card, which will yield the lowest possible latency.

Audio latency with WASAPI

So what kind of latencies can we get with WASAPI?

That depends on the device period that is being used. The term device period is a fancy way of saying buffer size; specifically the buffer size that is used in each call to your application that fetches audio data.

This is the same period with which audio data will be written out to the actual device, so it is the major contributor of latency in the entire machinery.

If you're using the AudioClient interface in WASAPI to initialize your streams, the default period is 10ms. This means the theoretical minimum latency you can get in shared mode would be 10ms (audio engine) + 10ms (driver) = 20ms. In practice, it'll be somewhat higher due to various inefficiencies in the subsystem.

When using exclusive mode, there's no engine latency, so the same number goes down to ~10ms.

These numbers are decent for most use-cases, but like I explained in my previous blog post, this is totally insufficient for pro-audio use-cases such as applying live effects to music recordings. You really need latencies that are lower than 10ms there.

Ultra-low latency with WASAPI

Starting with Windows 10, WASAPI removed most of its aforementioned inefficiencies, and introduced a new interface: AudioClient3. If you initialize your streams with this interface, and if your audio driver is implemented correctly, you can configure a device period of just 2.67ms at 48KHz.

The best part is that this is the period not just in exclusive mode but also in shared mode, which brings WASAPI almost at-par with JACK and CoreAudio

So that was the good news. Did I mention there's bad news too? Well, now you know.

The first bit is that these numbers are only achievable if you use Microsoft's implementation of the Intel HD Audio standard for consumer drivers. This is fine; you follow some badly-documented steps and it turns out fine.

Then you realize that if you want to use something more high-end than an Intel HD Audio sound card, unless you use one of the rare pro-audio interfaces that have drivers that use the new WaveRT driver model instead of the old WaveCyclic model, you still see 10ms device periods.

It seems the pro-audio industry made the decision to stick with ASIO since it already provides <5ms latency. They don't care that the API is proprietary, and that most applications can't actually use it because of that. All the apps that are used in the pro-audio world already work with it.

The strange part is that all this information is nowhere on the Internet and seems to lie solely in the minds of the Windows audio driver cabals across the US and Europe. It's surprising and frustrating for someone used to working in the open to see such counterproductive information asymmetry, and I'm not the only one.

This is where I plug open-source and talk about how Linux has had ultra-low latencies for years since all the audio drivers are open-source, follow the same ALSA driver model⁴, and are constantly improved. JACK is probably the most well-known low-latency audio engine in existence, and was born on Linux. People are even using Pulseaudio these days to work with <5ms latencies.

But this blog post is about Windows and WASAPI, so let's get back on track.

To be fair, Microsoft is not to blame here. Decades ago they made the decision of not working more closely with the companies that write drivers for their standard hardware components, and they're still paying the price for it. Blue screens of death were the most user-visible consequences, but the current audio situation is an indication that losing control of your platform has more dire consequences.

There is one more bit of bad news. In my testing, I wasn't able to get glitch-free capture of audio in the source element using the AudioClient3 interface at the minimum configurable latency in shared mode, even with critical thread priorities unless there was nothing else running on the machine.

As a result, this feature is disabled by default on the source element. This is unfortunate, but not a great loss since the same device period is achievable in exclusive mode without glitches.

Measuring WASAPI latencies

Now that we're back from our detour, the executive summary is that the GStreamer WASAPI source and sink elements now use the latest recommended WASAPI interfaces. You should test them out and see how well they work for you!

By default, a device is opened in shared mode with a conservative latency setting. To force the stream into the lowest latency possible, set low-latency=true. If you're on Windows 10 and want to force-enable/disable the use of the AudioClient3 interface, toggle the use-audioclient3 property.

To open a device in exclusive mode, set exclusive=true. This will ignore the low-latency and use-audioclient3 properties since they only apply to shared mode streams. When a device is opened in exclusive mode, the stream will always be configured for the lowest possible latency by WASAPI.

To measure the actual latency in each configuration, you can use the new audiolatency plugin that I wrote to get hard numbers for the total end-to-end latency including the latency added by the GStreamer audio ringbuffers in the source and sink elements, the WASAPI audio engine (capture and render), the audio driver, and so on.

I look forward to hearing what your numbers are on Windows 7, 8.1, and 10 in all these configurations! ;)


1. The only ones missing are AAudio because it's very new and ASIO which is a proprietary API with licensing requirements.

2. It's no secret that although lots of people use GStreamer on Windows, the majority of GStreamer developers work on Linux and macOS. As a result the Windows plugins haven't always gotten a lot of love. It doesn't help that building GStreamer on Windows can be a daunting task . This is actually one of the major reasons why we're moving to Meson, but I've already written about that elsewhere!

3. My knowledge about the history of the decisions behind the Windows Audio API is spotty, so corrections and expansions on this are most welcome!

4. The ALSA drivers in the Linux kernel should not be confused with the ALSA userspace library.

Latency in Digital Audio

We've come a long way since Alexander Graham Bell, and everything's turned digital.

Compared to analog audio, digital audio processing is extremely versatile, is much easier to design and implement than analog processing, and also adds effectively zero noise along the way. With rising computing power and dropping costs, every operating system has had drivers, engines, and libraries to record, process, playback, transmit, and store audio for over 20 years.

Today we'll talk about the some of the differences between analog and digital audio, and how the widespread use of digital audio adds a new challenge: latency.

Analog vs Digital

Analog data flows like water through an empty pipe. You open the tap, and the time it takes for the first drop of water to reach you is the latency. When analog audio is transmitted through, say, an RCA cable, the transmission happens at the speed of electricity and your latency is:

This number is ridiculously small—especially when compared to the speed of sound. An electrical signal takes 0.001 milliseconds to travel 300 metres (984 feet). Sound takes 874 milliseconds (almost a second).

All analog effects and filters obey similar equations. If you're using, say, an analog pedal with an electric guitar, the signal is transformed continuously by an electrical circuit, so the latency is a function of the wire length (plus capacitors/transistors/etc), and is almost always negligible.

Digital audio is transmitted in "packets" (buffers) of a particular size, like a bucket brigade, but at the speed of electricity. Since the real world is analog, this means to record audio, you must use an Analog-Digital Converter. The ADC quantizes the signal into digital measurements (samples), packs multiple samples into a buffer, and sends it forward. This means your latency is now:

We saw above that the first part is insignificant, what about the second part?

Latency is measured in time, but buffer size is measured in bytes. For 16-bit integer audio, each measurement (sample) is stored as a 16-bit integer, which is 2 bytes. That's the theoretical lower limit on the buffer size. The sample rate defines how often measurements are made, and these days, is usually 48KHz. This means each sample contains ~0.021ms of audio. To go lower, we need to increase the sample rate to 96KHz or 192KHz.

However, when general-purpose computers are involved, the buffer size is almost never lower than 32 bytes, and is usually 128 bytes or larger. For single-channel 16-bit integer audio at 48KHz, a 32 byte buffer is 0.33ms, and a 128 byte buffer is 1.33ms. This is our buffer size and hence the base latency while recording (or playing) digital audio.

Digital effects operate on individual buffers, and will add an additional amount of latency depending on the delay added by the CPU processing required by the effect. Such effects may also add latency if the algorithm used requires that, but that's the same with analog effects.

The Digital Age

So everyone's using digital. But isn't 1.33ms a lot of additional latency?

It might seem that way till you think about it in real-world terms. Sound travels less than half a meter (1½ feet) in that time, and that sort of delay is completely unnoticeable by humans—otherwise we'd notice people's lips moving before we heard their words.

In fact, 1.33ms is too small for the majority of audio applications!

To process such small buffer sizes, you'd have to wake the CPU up 750 times a second, just for audio. This is highly inefficient, and wastes a lot of power. You really don't want that on your phone or your laptop, and is completely unnecessary in most cases anyway.

For instance, your music player will usually use a buffer size of ~200ms, which is just 5 CPU wakeups per second. Note that this doesn't mean that you will hear sound 200ms after hitting "play". The audio player will just send 200ms of audio to the sound card at once, and playback will begin immediately.

Of course, you can't do that with live playback such as video calls—you can't "read-ahead" data you don't have. You'd have to invent a time machine first. As a result, apps that use real-time communication have to use smaller buffer sizes because that directly affects the latency of live playback.

That brings us back to efficiency. These apps also need to conserve power, and 1.33ms buffers are really wasteful. Most consumer apps that require low latency use 10-15ms buffers, and that's good enough for things like voice/video calling, video games, notification sounds, and so on.

Ultra Low Latency

There's one category left: musicians, sound engineers, and other folk that work in the pro-audio business. For them, 10ms of latency is much too high!

You usually can't notice a 10ms delay between an event and the sound for it, but when making music, you can hear it when two instruments are out-of-sync by 10ms or if the sound for an instrument you're playing is delayed. Instruments such as drum snare are more susceptible to this problem than others, which is why the stage monitors used in live concerts must not add any latency.

The standard in the music business is to use buffers that are 5ms or lower, down to the 0.33ms number that we talked about above.

Power consumption is absolutely no concern, and the real problems are the accumulation of small amounts of latencies everywhere in your stack, and ensuring that you're able to read buffers from the hardware or write buffers to the hardware fast enough.

Let's say you're using an app on your computer to apply digital effects to a guitar that you're playing. This involves capturing audio from the line-in port, sending it to the application for processing, and playing it from the sound card to your amp.

The latency while capturing and outputting audio are both multiples of the buffer size, so it adds up very quickly. The effects app itself will also add a variable amount of latency, and at 1.33ms buffer sizes you will find yourself quickly approaching a 10ms latency from line-in to amp-out. The only way to lower this is to use a smaller buffer size, which is precisely what pro-audio hardware and software enables.

The second problem is that of CPU scheduling. You need to ensure that the threads that are fetching/sending audio data to the hardware and processing the audio have the highest priority, so that nothing else will steal CPU-time away from them and cause glitching due to buffers arriving late.

This gets harder as you lower the buffer size because the audio stack has to do more work for each bit of audio. The fact that we're doing this on a general-purpose operating system makes it even harder, and requires implementing real-time scheduling features across several layers. But that's a story for another time!

I hope you found this dive into digital audio interesting! My next post will be is about my journey in implementing ultra low latency capture and render on Windows in the WASAPI plugin for GStreamer. This was already possible on Linux with the JACK GStreamer plugin and on macOS with the CoreAudio GStreamer plugin, so it will be interesting to see how the same problems are solved on Windows. Tune in!