Proximity voice chat: the built-in Sandbox.Voice component

s&box ships a complete, networked microphone/voice system. You do not need to hand-roll Steam voice capture, Opus transmission, or 3D playback. It is one engine component: Sandbox.Voice.

1. The component surface

Voice (public class Voice : Component) records and transmits voice/microphone input to other players. Attach one per player object. The full API surface:

Networking is built in. The component transmits via [Rpc.Broadcast(NetFlags.OwnerOnly | NetFlags.UnreliableNoDelay)]: the owner records (Steam voice, 44.1 kHz, ~30 Hz read tick), compresses, and broadcasts; receiving peers decompress and play through a SoundStream. IsListening returns false when IsProxy, so a peer only ever transmits its own character. Application.IsHeadless short-circuits playback. A dedicated server is silent; voice relays peer-to-peer via the broadcast. A static singleRecorder ensures only one Voice records per client.

2. The wiring recipe

Two rules carry the whole feature:

Rule A: attach to the networked player object BEFORE NetworkSpawn. The component rides the replicated snapshot to every peer (the component-ordering law: create components, then NetworkSpawn). Attach only in the networked path. A solo character gets no Voice, giving zero mic activity in single-player with no null-guards needed:

var go  = Scene.CreateObject();
var chr = go.Components.Create<MyCharacter>();
if ( owner is not null )
{
    go.Components.Create<Voice>();     // BEFORE NetworkSpawn
    go.NetworkSpawn( owner );
}

Rule B: configure on every peer, not via replication. Playback config on a proxy comes from that peer's own OnStart, not the snapshot. Set Mode, Distance, Falloff, and PushToTalkInput on every peer:

protected override void OnStart()
{
    var voice = Components.Get<Voice>();
    if ( !voice.IsValid() ) return;   // solo mode - no Voice attached

    voice.Mode           = Voice.ActivateMode.PushToTalk;
    voice.PushToTalkInput = "Voice";
    voice.WorldspacePlayback = true;
    voice.LipSync        = true;
    voice.Distance       = VoiceMaxRangeM * 39.37f;   // metres → s&box units

    float nearRatio = VoiceNearRangeM / VoiceMaxRangeM;
    voice.Falloff = new Curve(
        new Curve.Frame( 0f,        1f, 0f, 0f ),
        new Curve.Frame( nearRatio, 1f, 0f, 0f ),
        new Curve.Frame( 1f,        0f, 0f, 0f ) );

    voice.Renderer = Components.Get<SkinnedModelRenderer>();
}

Curve(params Frame[]) and Curve.Frame(time, value, inTangent, outTangent) are public. Do not ship the default Falloff + Distance for proximity chat: 381 m of range with the stock curve gives a mushy, map-spanning whisper instead of a crisp local bubble.

3. Push-to-talk: three layers

1. The engine user preference is already PTT by default. Preferences.VoiceMode (ConVar voip_mode, saved) defaults to PushToTalk. The IsListening logic: Disabled never transmits; PushToTalk always requires Input.Down(PushToTalkInput) regardless of component Mode; OpenMicrophone honors the component Mode.
2. Bind the built-in "Voice" input action. Every project's ProjectSettings/Input.config typically ships a "Voice" action (often with KeyboardCode: None); the component's default PushToTalkInput = "voice" binds to it. Assign a free key and you're done. Verify it exists: not every project has it. If missing, add it to Input.config (editor restart required for input config changes to register).
3. Set component Mode = PushToTalk anyway (belt-and-suspenders): keeps PTT gating even for a user whose saved voip_mode is OpenMicrophone. Without it, that user would hot-mic your game.

4. The speaking-indicator pattern

On a proxy, voice.LastPlayed < holdSeconds is a flicker-free "speaking" flag: packets arrive ~30/s while the mic is active, so a ~0.3 s hold window bridges inter-packet gaps and clears shortly after the talker stops:

var voice = chr.Components.Get<Voice>();
bool speaking = voice.IsValid() && voice.LastPlayed < 0.3f;

Anchor the tag off the networked transform, not an owner-only sim field. A proxy (the only character you tag) never runs the owner's OnFixedUpdate, so any private position field written there (and not [Sync]) freezes at the proxy's spawn value while the replicated GameObject.WorldPosition carries the real player. Anchor the tag off a frozen field and it strands at spawn while the model walks away. Read the character's WorldPosition (the same interpolated source the parented visual renders from). It also satisfies the read-interpolated-transforms rule. In a live two-peer session, a tag anchored to an unsynced owner field was observed hanging over the spawn area while the player model stood elsewhere; switching to WorldPosition fixed it immediately.

For a Razor name-tag overlay, fold the flag into BuildHash so the DOM rebuilds only on a state change, not per frame. Components.Get<Voice>() returns null on a non-networked character, so speaking = false gives the correct solo behavior for free.

5. Falloff radii as tuning constants

House the radii as named constants in metres with doc comments; convert once at emission (s&box units are inches; 1 m = 39.37 units):

public const float VoiceNearRangeM    = 8f;    // full-volume conversational core
public const float VoiceMaxRangeM     = 35f;   // silent at/beyond
public const float VoiceSpeakingHoldS = 0.3f;  // indicator hold window

These are live-tuning dials: the "how far can I hear them" feel can only be judged in a real two-peer session. Scale to your world footprint; the engine default (381 m) is effectively "everywhere" on a small map.

6. Extension points

• Per-player mute: subclass (class MyVoice : Voice) overriding ExcludeFilter() to yield the connections that should not hear you (applied via Rpc.FilterExclude around the broadcast).
• Per-player deafen: override ShouldHearVoice(Connection) to return false for speakers the local user muted.
• Lip-sync: set Renderer to the peer's SkinnedModelRenderer. The engine drives citizen viseme morphs from the live analysis, holds the shape through packet gaps, and eases out after ~1 s. If the model lacks viseme morphs, LipSync safely no-ops.
• Team voice / screen-space stereo: WorldspacePlayback = false puts the sound in screen space for team-left/right audio.
• Occlusion: enabled by default in worldspace playback. Only avoidable via a subclass override if terrain-muffled voice is wrong for your game.
• Mixer routing: VoiceMixer targets any descendant of the Voice mixer, the hook for a user voice-volume slider (don't repurpose Volume for that).

7. The two-peer verification checklist

Run once per project after wiring. Each row should be proven live:

Check #2 is the keystone: it proves the create-before-NetworkSpawn premise (that a proxy's Msg_Voice RPC arrives and plays). Everything upstream of #2 is compile-proven only.

8. Pitfalls