Adding Mumble Support to Valheim

21 Feb 2021

I'm taking a detour from my usual Rust side projects to dabble in something new: Unity modding! I, like many others, have found myself addicted obsessed highly invested in Valheim, a multiplayer viking survival game that recently entered Steam Early Access. Lacking any built-in voice capabilities, players use other platforms to communicate. Mumble, an open-source voice chat program, has a cool feature where it can use in-game position data to make it sound like your friends' voices actually come from their characters. The trick is getting that data into Mumble.


Because nothing exists in a vaccuum and everything we build is on the shoulders of giants, I'll first go over the basic modding setup and tools for exploration. Valheim is built with Unity and doesn't go though the IL2CPP converter, so pretty much all of its code ends up as a C# assembly. This makes it fairly easy to open up with a tool like ILSpy to decompile it and view the source. That helps understand how the game works, but what about actually adding code to it?

That's where BepInEx comes in. BepInEx is a modding framework and loader for Unity games that use C# for scripting. Installation is extremely simple - basically just dropping its files into the game installation directory. The only issue with this is that Valheim ships with stripped Unity assemblies. They've had unneeded functionality removed, so they only contain what the game actually uses. BepInEx, unfortunately, requries some of that removed functionality, and it crashes immediately when installed. Luckily, some kind soul has put together a package that includes the un-stripped Unity assemblies.

Lastly, there are a couple of utility mods that make the the process building a new one a lot simpler. UnityExplorer provides a handy interface for browsing and poking at Unity GameObjects (and some other stuff) at runtime. ScriptEngine lets you reload plugins without having to restart the game ever time you recompile it.

Before diving into more Unity stuff, we need to understand what Mumble needs to do its thing. If you're only here for the Unity stuff, this can be safely skipped. Documentation is a bit sparse, but Mumble Wiki pages on Positional Audio and Link have most of what's needed to get it all connected. To boil down the "linking" process to its basics: the game and mumble use a shared memory-mapped file with a layout that's understood by both. The game updates this shared memory with location data every frame, and mumble reads it fifty times per second. The shared memory has fields for character and camera locations and headings, the game name, a player identifier, a "context", and a "tick" counter, so we'll need a way to get all of these values from the game. Most fields seem fairly self-explanatory, but context warrants a bit more. If two players have the same context, they'll get directional audio from each other. If they have different contexts, They'll just get normal non-directional audio. This will be important later.

Platform Differences

Linking a game to Mumble doesn't require any Mumble-specific libraries - it's all done via standard OS functionality. Unfortunately, this means that the exact mechanism differes slightly between Unix-like systems and Windows.

The Mumble Link example code for Windows uses the File Mapping C API:

HANDLE hMapObject = OpenFileMappingW(FILE_MAP_ALL_ACCESS, FALSE, L"MumbleLink");
lm = (LinkedMem *) MapViewOfFile(hMapObject, FILE_MAP_ALL_ACCESS, 0, 0, sizeof(LinkedMem));

The Unix-like example code uses the POSIX shared memory API:

char memname[256];
snprintf(memname, 256, "/MumbleLink.%d", getuid());

int shmfd = shm_open(memname, O_RDWR, S_IRUSR | S_IWUSR);
lm = (LinkedMem *)(mmap(NULL,
                        sizeof(struct LinkedMem),
                        PROT_READ | PROT_WRITE,

The end result is the same: lm points to a region of memory that's the size of LinkedMem and can be shared by the game and Mumble processes.

The example code is in C, but our plugin is going to need to be in C#, so some porting is going to be required. C#, being extremely Windows-y, makes the Windows case quite straightforward:

memoryMappedFile = MemoryMappedFile.CreateOrOpen(
byte* tmp = null;
    .AcquirePointer(ref tmp);
ptr = (WindowsLinkMemory*)tmp;

For the POSIX case, it's a bit less straightforward. There aren't any built-in ways to call the necessary functions that I could find, so the easiest approach seemed to be to call libc directly. Fortunately, C# makes this quite easy:

// In a class definition somewhere.
private static extern int shm_open([MarshalAs(UnmanagedType.LPStr)] string name, int oflag, uint mode);
private static extern uint getuid();
private static extern void* mmap(void* addr, long length, int prot, int flags, int fd, long off);

// In the constructor
fd = shm_open("/MumbleLink." + getuid(), 2, 384);
ptr = (LinuxLinkMemory*)mmap(null, Marshal.SizeOf<LinuxLinkMemory>(), 3, 1, fd, 0);

DllImport allows you to declare external functions along with the shared library they're defined in. All loading of the shared library happens at runtime, so it pretty much "just works," with the added benefit of not throwing linking errors on Windows as long as you don't try to call the methods. One thing I found really cool is the MarshalAs tag on the first argument to shm_open. This is an attribute that can be placed on extern function arguments that tells the runtime how to convert the type in the declaration to what the function actually expects. In this case, it takes our string and turns it into a null-terminated char* in the usual POSIX fashion.

Memory Layout

When the Mumble Link memory layout was defined, a very unfortunate decision was made. The C version is defined like so:

struct LinkedMem {
#ifdef _WIN32
	UINT32	uiVersion;
	DWORD	uiTick;
	uint32_t uiVersion;
	uint32_t uiTick;
	float	fAvatarPosition[3];
	float	fAvatarFront[3];
	float	fAvatarTop[3];
	wchar_t	name[256];
	float	fCameraPosition[3];
	float	fCameraFront[3];
	float	fCameraTop[3];
	wchar_t	identity[256];
#ifdef _WIN32
	UINT32	context_len;
	uint32_t context_len;
	unsigned char context[256];
	wchar_t description[2048];

We're just going to ignore the odd DWORD there. The Microsoft Data Types Specification defines it as a 32-bit integer, so it's exactly the same size as the UINT32's that are used elsewhere. The real egregious type choice is wchar_t. Special care was taken to ensure that every other type had the same width regardless of platform/compiler, except the fields that are strings. The CPP reference states that wchar_t is "Required to be large enough to represent any supported character code point." On most systems, this means that it's an unsigned 32-bit integer and can simply contain the un-encoded unicode codepoint. On Windows, however, they're half that size - 16 bits, and contain UTF-16-encoded characters. This means that the LinkMemory has a different layout on different systems, and some extra special care has to be taken in modern languages that assumes that types are consistent across platforms. On Linux, this means that the string fields are fixed uint field[len], while on Windows, they're fixed ushort field[len]. The encoding used to write them also needs to be UTF32 and Unicode respectively.

I'm not sure why the strings were laid out this way. My best guess is that it's from the days when pretty much everything in Windows was UTF-16, and using wchar_t and wcsncpy kept the C interface consistent. It might have made made sense when everything was written in C, but it only serves to make my life more difficult today.

Putting It Together

Actually switching between the two implementations is pretty boring after the memory layout was all figured out. There are two options: C-style #if preprocessor directives that are evaluated at compile-time, or including both versions in the assembly and switching between them at runtime. To avoid having to compile and distribute both versions, I opted for the later. The end result is a couple of interfaces with implementations for both Windows and Linux:

public interface LinkFileFactory
    LinkFile Open();

public interface LinkFile : IDisposable
    uint UIVersion { set; }
    void Tick();
    Vector3 CharacterPosition { set; }
    Vector3 CharacterForward { set; }
    Vector3 CharacterTop { set; }
    string Name { set; }
    Vector3 CameraPosition { set; }
    Vector3 CameraForward { set; }
    Vector3 CameraTop { set; }
    string ID { set; }
    string Context { set; }
    string Description { set; }

Back to Unity

Now that the Mumble stuff is out of the way, all there is to do is to find the data that we need, add it to the LinkFile, and increment its tick counter. Easy, right?

Objects of Interest

For best results, we need both the character location and camera location. Players expect to hear things in relation to the camera, regardless of how the character is oriented, and expect to hear things coming from other players' character rather than their orbiting camera. Thus, both positions and headings are needed. We don't need the *Top vectors. Those are only needed if you can tilt your head side to side in-game, like in first-person shooters where you can lean around corners and such.

I did most of my investigation in UnityExplorer. After loading a character into a world, there's a big obvious _GameMain object at the root of the main scene. Seems promising, right? Inside _GameMain is a Main Camera, score! The Main Camera object has a child named Listner. Its components are extremely telling of their purpose.

Main Camera

Any game that has directional audio, regardless of voice chat, has to already have some object with the exact position and heading information needed to figure out where sounds are coming from. For Valheim, Listner is that object, so it only makes sense to reuse it for our purposes. You can even inspect its Transform (Unity speak for location info), check "Auto Update," and watch as the forward and position vectors change as you move the camera around. Pretty neat!

The character information was a little trickier to find. I poked around in _GameMain quite a bit without much luck, and explored some of the other objects as well. _NetSceneRoot has gobs of children, most or all of which seem to correspond to physical objects in the game world. Unfortunately, there are over 200 pages of them to scroll through, so just stumbling across the player object probably won't work. Luckily, UnityExplorer has a "Mouse Inspector" mode. At the right angle, you can inspect the player object

Inspect Player Player Object


With the Listner and Player objects, we've got all of the location information needed to feed to Mumble. So what about the things like the context? On one hand, we could just set it to a static value and everyone playing should get directional audio from each other. But what if some people are connected to a different server? Putting aside how generally confusing that would be, it would be even more confusing if you were hearing the "ghost" of someone else directionally who wasn't even in your world. So having separate contexts per world seems the way to go.

This ended up being more difficult to find than the player and camera information. Eventually, after decompiling and grep'ing around the source, I landed upon the ZNet class. This seems to be the connection manager for the game, and is running even in singleplayer mode or as the host of a game. It has a method to GetWorldName, which sounds perfect for use as a context string.

As an added bonus, using the world name as the context allowed me to test the mod with steam in offline mode, so I didn't have to recruit a friend to babble at. By loading into the same world on different single-player sessions, both used the same context, so I could at least verify that I was in fact getting positional sound, even if I couldn't actually see the source of it. It does have the downside of this being a possibility in the event that two Mumble users are connected to two different servers with the same world name, but that seems like an acceptable edge case to me.

The Mod Itself

The basic project structure is pretty well covered by the BepInEx Tutorial, so I won't rehash it here. The "meat" of it is how and when it decides to link to Mumble. C# scripts in Unity implement the MonoBehavior interface. They have a number of methods that are called at various points in their lifetime. The ones we care about are Awake and FixedUpdate.

Awake runs just after the object is instantiated. In the case of BepInEx plugins, this is close to when the game starts. I use it to figure out the platform that it's running on, and select the proper Mumble Link implementation.

void Awake()
    if (Environment.OSVersion.Platform == PlatformID.Win32NT)
        UnityEngine.Debug.Log("Windows Mumble Link Loaded");
        fileManager = new mumblelib.WindowsLinkFileManager();
        UnityEngine.Debug.Log("Unix Mumble Link Loaded");
        fileManager = new mumblelib.UnixLinkFileManager();

FixedUpdate runs on a fixed interval (shocker, right?) and is intended to be used for things like physics calculations that aren't framerate-dependent. This makes it the perfect choice to gather our location data and send it off to Mumble.

unsafe void FixedUpdate()
    // findGameObjects attempts to find the player and camera objects if they
    // haven't already been discovered. If they have been, it verifies that
    // they still have valid transforms.
    if (!findGameObjects())
        // If Mumble was previously linked, unlink it.
        if (mumbleLink != null)
            mumbleLink = null;

    // If mumble hasn't yet been linked, link and initialize the static values.
    if (mumbleLink == null)

    Transform camera = cameraObject.transform;
    Transform character = characterObject.transform;

    if (camera.position != null)
        mumbleLink.CameraPosition = camera.position;
    if (camera.forward != null)
        mumbleLink.CameraForward = camera.forward;
    if (character.position != null)
        mumbleLink.CharacterPosition = character.position;
    if (character.forward != null)
        mumbleLink.CharacterForward = character.forward;


Locating the GameObjects is an inefficient process, so we only want to do it once, and then keep them around until they're no longer valid. findGameObjects handles all of this for us:

private bool findGameObjects()
    // Only look for the player and camera if they haven't been found previously.
    if (!inGame)
        characterObject = GameObject.Find("_NetSceneRoot/Player(Clone)");
        cameraObject = GameObject.Find("_GameMain/Main Camera/Listner");

    if ((!isValid(characterObject)) ||
        (!isValid(cameraObject)) ||
        (ZNet.instance == null))
        if (inGame)
            UnityEngine.Debug.Log("Lost game objects");
            inGame = false;
        return false;

    if (!inGame)
        UnityEngine.Debug.Log("Found game objects");
        inGame = true;
    return true;

private static bool isValid(GameObject gameObject)
    return gameObject != null && gameObject.transform != null;

Init is pretty boring but I'll include it just for completeness' sake.

unsafe void Init()
    mumbleLink = fileManager.Open();
    mumbleLink.Name = "Valheim";
    mumbleLink.Description = "Valheim Positional Audio";
    mumbleLink.UIVersion = 2;
    string id = randomString(16);
    UnityEngine.Debug.LogFormat("Setting Mumble ID to {0}", id);
    mumbleLink.ID = id;
    UnityEngine.Debug.LogFormat("Setting context to {0}", ZNet.instance.GetWorldName());
    context = ZNet.instance.GetWorldName();
    mumbleLink.Context = context;

    UnityEngine.Debug.Log("Mumble Shared Memory Initialized");

Bonus for the Modern Viking

Some may enjoy playing with Mumble's positional audio set up such that they can't hear people who are too far away. But what about when they need to talk to someone halfway across the world? A feature like Phasmophobia's walkie-talkie would be neat. In Phasmophobia, voice chat is mandatory, and is all positional by default. You have a button you can press, however, to speak to everyone in the game, regardless of where you are in relation to each other. This is easy enough to simulate with some clever manipulation of the context field.

Recall that two players with the same context will hear each other positionally, but if they have different contexts, they'll get "normal" audio from each other. We can therefore keep track of the original context, and set it to gibberish while a key is held down, thus making your context different from everyone else's and temporarily disabling positional audio. It also lets you listen in on everyone else independent of positions.

The easiest way to accomplish this is with the KeyboardShortcut configuration option for BepInEx. By adding this to the Awake method, a configuration file gets generated, and mods like BepInEx.ConfigurationManager can edit the setting.

globalVoice = Config.Bind(
    new BepInEx.Configuration.KeyboardShortcut(KeyCode.T),
    "Global voice key"

It can then be checked in the Update method like so:

void Update()
    if (mumbleLink != null)
        if (globalVoice.Value.IsPressed())

private void enterGlobalContext()
    if (!inGlobal)
        // A gibberish context will make your voice non-positional to everyone else.
        UnityEngine.Debug.Log("Switching to global voice context");
        mumbleLink.Context = randomString(16);
        inGlobal = true;

private void exitGlobalContext()
    if (inGlobal)
        UnityEngine.Debug.Log("Switching to local voice context");
        mumbleLink.Context = context;
        inGlobal = false;

There's likely a better way to accomplish the viking walkie-talkie effect via server-side scripts and clever use of groups and ACLs, but this approach is minimally invasive and doesn't require any additional configuration.


Full source for the plugin can be found here, and the compiled version can be downloaded from the Releases. Bug reports are always welcome, as are suggestions for improvement!