Capture every audio stream into its own ring and mix them on the host

Games with several concurrent WASAPI render streams (e.g. Spider-Man: Miles Morales) only had their first ("primary") stream mirrored; the rest kept playing locally and never reached the guest. Now the render-hook captures + silences EVERY tracked stream into its own ring (coop_audio_<pid>[_<index>]), each published with that stream's own detected format (Initialize when caught, else GetMixFormat -- the per-stream format detection, now actually used per ring rather than only for the primary). The host creates a ring per stream and mixes the same-format streams with a soft clip (host/src/audio/audio_mix.hpp); streams whose format differs from the primary are still silenced (no echo) but skipped from the mix (would need resampling). The single-stream case is byte-for-byte unchanged: when only one stream is active the host passes it through without the mixer, so the common path has no overhead or fidelity change. Verified: new audio_mix_test covers the decode/sum/soft-clip/encode math (float32 + int16); audio_hook_test (x64 + x86) still passes, guarding the primary capture+silence path against regression; full build x64 + x86 clean; ctest x64 11/11, x86 3/3. Multi-stream mixing against a real multi-stream game needs a live session to fully confirm. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-21 05:45:47 +02:00
parent dfd2be8c17
commit 784c31a9b5
10 changed files with 412 additions and 147 deletions
--- a/README.md
+++ b/README.md
@@ -29,7 +29,7 @@ and forwards guest controllers back into it.
 | Keep game running unfocused | Hook spoofs focus so the game polls while the host holds OS focus | `coop_hook.dll` |
 | Mirror video (default) | Windows Graphics Capture of the game window, letterboxed into the host window | `coop_host.exe` |
 | Mirror video (hooked) | Injected Present / OpenGL hook copies the backbuffer into a shared keyed-mutex texture the host samples (lower latency, no capture border) | `coop_hook.dll` + `coop_host.exe` |
-| Mirror audio | Injected render-hook copies the game's WASAPI frames into a shared ring and silences the game locally (no echo); WASAPI process loopback is the automatic fallback | `coop_hook.dll` + `coop_host.exe` |
+| Mirror audio | Injected render-hook copies each of the game's WASAPI render streams into its own shared ring and silences the game locally (no echo); the host mixes the streams (soft-clipped); WASAPI process loopback is the automatic fallback | `coop_hook.dll` + `coop_host.exe` |
 | Host ↔ hook IPC | Named shared memory (seqlock for input, status back-channel, video/audio/log shares) | `common/` |

 The hooked video path has two producers: **Direct3D (DXGI)** hooks
@@ -72,23 +72,12 @@ and covers anything the hooked path doesn't (Vulkan, D3D9 — see Roadmap).

 ### Planned (next up)

-Worked top-to-bottom: each item is a milestone with its own tests and commit, and
-is removed from this list once done — so the top item is always next. The
-self-verifiable tooling / UI / input items come first; the game-pipeline items that
-need a real game (and Remote Play) to fully validate come last.
-
- **Multi-stream audio capture + mixing, with per-stream format detection.** Games
-  with several concurrent WASAPI render streams (e.g. Miles Morales) only get their
-  first ("primary") stream mirrored today; the rest keep playing locally and never
-  reach the guest. Capture every tracked render stream into its own shared ring,
-  silence each, and add a host-side mixer that resamples each ring to the render
-  format and sums them (with soft-clip). **Fold in real per-stream format
-  detection** here, since it touches the same hook + ring plumbing: a stream that
-  already existed when we injected is never seen at `Initialize`, so the hook
-  currently assumes the device **mix format** and a shared-mode stream opened at a
-  different format comes out wrong-pitched. Resolve each stream's true format (the
-  stream's own `Initialize` when caught, else the original `GetMixFormat`) so every
-  mixed ring is pitched correctly.
+Nothing queued — the previous backlog (bin restructure, terminated/hung detection,
+re-attach, window-based target picker, overlay auto-layout, Audio "live" column,
+moving the synthetic-input toggle, mouse + keyboard forwarding, rumble forwarding,
+per-backend input debug view, cursor release, capture metrics + latency, DX12 hooked
+capture, multi-stream audio + per-stream formats) is all shipped. See Future work
+for what's left.

 ### Future work

@@ -152,6 +141,8 @@ ctest --test-dir build -C Debug --output-on-failure
  no controller needed).
 - **`audio_ring_test`** — unit test of the shared audio ring (lock-free SPSC
  push/pop, wrap-around, format handshake, overrun/drop). No device needed.
+- **`audio_mix_test`** — unit test of the multi-stream mixer math (decode / sum /
+  soft-clip / encode for float32 + int16). No device needed.
 - **`audio_hook_test`** — in-process self-test of the WASAPI render-hook: installs
  the hooks, renders a tone through WASAPI in the same process, and asserts the
  COM vtables were discovered, the frames reached the ring (non-silent), the
--- a/common/include/coop/audio_ring.hpp
+++ b/common/include/coop/audio_ring.hpp
@@ -185,10 +185,18 @@ inline std::uint32_t audio_ring_pop(AudioRingHeader& h, void* dst, std::uint32_t
 	return bytes;
 }

-// Build the per-pid audio ring name both sides agree on.
-inline std::wstring audio_ring_name(unsigned long target_pid)
+// Build the per-pid audio ring name both sides agree on. Stream 0 keeps the bare
+// coop_audio_<pid> name (backward compatible / the single-stream case); additional
+// streams append _<index> (coop_audio_<pid>_1, _2, ...). The host captures every
+// render stream into its own ring and mixes them.
+inline std::wstring audio_ring_name(unsigned long target_pid, unsigned index = 0)
 {
-	return std::wstring(kAudioRingPrefix) + std::to_wstring(target_pid);
+	std::wstring name = std::wstring(kAudioRingPrefix) + std::to_wstring(target_pid);
+	if (index != 0)
+	{
+		name += L"_" + std::to_wstring(index);
+	}
+	return name;
 }

 } // namespace coop
--- a/hook/src/audio_hook.cpp
+++ b/hook/src/audio_hook.cpp
@@ -113,7 +113,9 @@ struct CapturedFormat
 // --- Global hook state -----------------------------------------------------

 IpcClient* g_ipc = nullptr;
-std::atomic<AudioRingHeader*> g_ring{nullptr};
+// One ring per tracked stream (index = the stream's debug slot). Stream 0 is the
+// primary; the host creates a ring per stream and mixes them.
+std::atomic<AudioRingHeader*> g_rings[kMaxAudioStreams]{};

 std::mutex g_setup_mutex; // guards installs + the format map + stream registration

@@ -146,32 +148,28 @@ std::atomic<IAudioRenderClient*> g_self_render{nullptr};
 CapturedFormat g_mix_format;
 std::atomic<std::uint32_t> g_have_mix_format{0};

-// The primary stream's actual format, captured when it's registered. The host
-// creates the ring only when audio mirroring is toggled on — typically *after*
-// the primary stream was already registered — so the format must be (re)published
-// to the ring whenever it attaches. (The in-process probe creates the ring before
-// injecting, so it never exercises this ordering; the full app always does.)
-// Guarded by g_setup_mutex.
-CapturedFormat g_primary_format;
+// Each tracked stream's actual format, captured when it's registered. The host
+// creates the rings only when audio mirroring is toggled on — typically *after* a
+// stream was already registered — so the format must be (re)published to a ring
+// whenever it attaches. Guarded by g_setup_mutex.
+CapturedFormat g_stream_formats[kMaxAudioStreams];

 // Per IAudioClient, the format captured at Initialize, looked up when its render
 // client is created. Setup-path only (never touched on the audio thread).
 std::unordered_map<IAudioClient*, CapturedFormat> g_client_formats;

-// Streams we track for the debug view (frame counting). Index 0 is primary.
+// Streams we track (frame counting + per-stream capture). Index 0 is primary.
 struct TrackedStream
 {
 	std::atomic<IAudioRenderClient*> client{nullptr};
 	std::atomic<std::uint64_t> frames{0};
+	std::atomic<std::uint32_t> block_align{0}; // hot-path frame size for this stream
 };
 TrackedStream g_streams[kMaxAudioStreams];
 std::uint32_t g_registered = 0;            // slots filled (<= kMaxAudioStreams), under mutex
 std::atomic<std::uint32_t> g_streams_seen{0}; // total distinct clients ever seen

-// Hot-path primary identity + frame size (avoids touching the map/mutex).
-std::atomic<IAudioRenderClient*> g_primary{nullptr};
-std::atomic<std::uint32_t> g_primary_block_align{0};
-std::atomic<std::uint64_t> g_frames_captured{0};
+std::atomic<std::uint64_t> g_frames_captured{0}; // total frames captured across streams

 // GetBuffer/ReleaseBuffer are paired on one thread, never nested: stash the
 // pointer the game just got so ReleaseBuffer can copy it before releasing.
@@ -235,35 +233,34 @@ HRESULT STDMETHODCALLTYPE hk_ReleaseBuffer(IAudioRenderClient* self, UINT32 num_
 		try_register_lazy(self);
 	}

-	// Frame counting for any tracked stream (drives the live/idle debug view).
+	// Per tracked stream: count frames (debug view) and, into the stream's own ring,
+	// capture + silence its buffer while capture is enabled. Every stream is captured
+	// into its own ring; the host mixes them.
 	for (std::uint32_t i = 0; i < kMaxAudioStreams; ++i)
 	{
-		if (g_streams[i].client.load(std::memory_order_acquire) == self)
+		if (g_streams[i].client.load(std::memory_order_acquire) != self)
 		{
+			continue;
+		}
 		const std::uint64_t total =
 		    g_streams[i].frames.fetch_add(num_frames, std::memory_order_relaxed) + num_frames;
 		if (g_ipc != nullptr)
 		{
 			g_ipc->note_audio_frames(i, total);
 		}
-			break;
-		}
-	}

-	// Capture + silence only the primary stream, only while enabled.
-	if (self == g_primary.load(std::memory_order_acquire) && num_frames > 0 &&
-	    (flags & AUDCLNT_BUFFERFLAGS_SILENT) == 0)
+		if (num_frames > 0 && (flags & AUDCLNT_BUFFERFLAGS_SILENT) == 0)
 		{
-		AudioRingHeader* ring = g_ring.load(std::memory_order_acquire);
+			AudioRingHeader* ring = g_rings[i].load(std::memory_order_acquire);
 			if (ring != nullptr && ring->capture_enabled.load(std::memory_order_relaxed) != 0 &&
 			    t_gb_client == self && t_gb_data != nullptr && t_gb_frames == num_frames)
 			{
-			const std::uint32_t block = g_primary_block_align.load(std::memory_order_relaxed);
+				const std::uint32_t block = g_streams[i].block_align.load(std::memory_order_relaxed);
 				const std::uint32_t bytes = num_frames * block;
 				// Only silence if the frames made it into the ring; if the host has
 				// stalled (ring full) keep playing locally rather than going dead
 				// silent — degrades to today's echo, never to silence.
-			if (audio_ring_push(*ring, t_gb_data, bytes, num_frames))
+				if (block != 0 && audio_ring_push(*ring, t_gb_data, bytes, num_frames))
 				{
 					std::memset(t_gb_data, 0, bytes); // belt-and-suspenders vs a driver ignoring SILENT
 					g_frames_captured.fetch_add(num_frames, std::memory_order_relaxed);
@@ -272,6 +269,8 @@ HRESULT STDMETHODCALLTYPE hk_ReleaseBuffer(IAudioRenderClient* self, UINT32 num_
 				}
 			}
 		}
+		break;
+	}
 	return g_vh_releasebuffer.original<ReleaseBufferFn>()(self, num_frames, flags);
 }

@@ -303,7 +302,9 @@ void register_render_client_locked(IAudioRenderClient* rc, const CapturedFormat&
 	}
 	g_registered = slot + 1;

+	g_stream_formats[slot] = cf;
 	g_streams[slot].frames.store(0, std::memory_order_relaxed);
+	g_streams[slot].block_align.store(cf.block_align, std::memory_order_relaxed); // before client (hot path)
 	g_streams[slot].client.store(rc, std::memory_order_release);

 	AudioStreamInfo info{};
@@ -318,19 +319,14 @@ void register_render_client_locked(IAudioRenderClient* rc, const CapturedFormat&
 		g_ipc->publish_audio_stream(slot, info);
 	}

-	if (slot == 0)
-	{
-		g_primary_format = cf;
-		g_primary_block_align.store(cf.block_align, std::memory_order_relaxed);
-		g_primary.store(rc, std::memory_order_release);
-		AudioRingHeader* ring = g_ring.load(std::memory_order_acquire);
-		logf("primary stream set: rc=%p ring=%p (format %s)", rc, ring,
+	// Publish this stream's format to its own ring if the host has attached one yet.
+	AudioRingHeader* ring = g_rings[slot].load(std::memory_order_acquire);
+	logf("stream %u set: rc=%p ring=%p fmt=%uHz/%uch/%ubit (%s)", slot, rc, ring, cf.rate, cf.channels, cf.bits,
 	     ring ? "published" : "no ring yet");
-		if (ring)
+	if (ring != nullptr)
 	{
 		audio_ring_set_format(*ring, cf.rate, cf.channels, cf.bits, cf.tag, cf.block_align);
 	}
-	}
 	// GetBuffer/ReleaseBuffer are hooked proactively at install time (the shared
 	// vtable covers every render client), so nothing to install per-stream here.
 }
@@ -464,7 +460,7 @@ bool install_audio_hooks(IpcClient& ipc, AudioRingHeader* ring)
 {
 	std::scoped_lock lock(g_setup_mutex);
 	g_ipc = &ipc;
-	g_ring.store(ring, std::memory_order_release);
+	g_rings[0].store(ring, std::memory_order_release);
 	if (g_vh_activate)
 	{
 		return true; // anchor already installed
@@ -568,33 +564,48 @@ bool install_audio_hooks(IpcClient& ipc, AudioRingHeader* ring)

 void republish_audio_format()
 {
-	AudioRingHeader* ring = g_ring.load(std::memory_order_acquire);
-	if (ring == nullptr || audio_ring_format_ready(*ring))
+	// Nothing to do if no ring needs a format yet (cheap pre-check, no lock).
+	bool any_pending = false;
+	for (std::uint32_t i = 0; i < kMaxAudioStreams; ++i)
 	{
-		return; // no ring yet, or the format is already published
+		AudioRingHeader* ring = g_rings[i].load(std::memory_order_acquire);
+		if (ring != nullptr && !audio_ring_format_ready(*ring))
+		{
+			any_pending = true;
+			break;
+		}
+	}
+	if (!any_pending)
+	{
+		return;
 	}
 	std::scoped_lock lock(g_setup_mutex);
-	if (audio_ring_format_ready(*ring))
+	for (std::uint32_t i = 0; i < kMaxAudioStreams; ++i)
 	{
-		return; // raced with another publisher
+		AudioRingHeader* ring = g_rings[i].load(std::memory_order_acquire);
+		if (ring == nullptr || audio_ring_format_ready(*ring) || g_stream_formats[i].rate == 0)
+		{
+			continue; // no ring, already published, or this slot has no stream yet
 		}
-	if (g_primary.load(std::memory_order_acquire) != nullptr && g_primary_format.rate != 0)
-	{
-		audio_ring_set_format(*ring, g_primary_format.rate, g_primary_format.channels, g_primary_format.bits,
-		                      g_primary_format.tag, g_primary_format.block_align);
-		logf("republish_audio_format: published %uHz/%uch/%ubit to ring %p", g_primary_format.rate,
-		     g_primary_format.channels, g_primary_format.bits, ring);
+		const CapturedFormat& cf = g_stream_formats[i];
+		audio_ring_set_format(*ring, cf.rate, cf.channels, cf.bits, cf.tag, cf.block_align);
+		logf("republish_audio_format: stream %u -> %uHz/%uch/%ubit ring %p", i, cf.rate, cf.channels, cf.bits,
+		     ring);
 	}
 }

-void set_audio_ring(AudioRingHeader* ring)
+void set_audio_ring(unsigned index, AudioRingHeader* ring)
 {
-	g_ring.store(ring, std::memory_order_release);
-	logf("set_audio_ring: ring=%p capture_enabled=%u", ring,
+	if (index >= kMaxAudioStreams)
+	{
+		return;
+	}
+	g_rings[index].store(ring, std::memory_order_release);
+	logf("set_audio_ring: index=%u ring=%p capture_enabled=%u", index, ring,
 	     ring ? ring->capture_enabled.load(std::memory_order_relaxed) : 0u);
-	// The primary may already be registered (game was playing before we injected
-	// and before the host created the ring); publish its format so the host stops
-	// waiting and consumes the ring instead of falling back to loopback.
+	// The stream may already be registered (game was playing before we injected and
+	// before the host created the ring); publish its format so the host stops waiting
+	// and consumes the ring instead of falling back to loopback.
 	republish_audio_format();
 }

@@ -624,20 +635,19 @@ void remove_audio_hooks()
 		g_self_client = nullptr;
 	}
 	g_have_mix_format.store(0, std::memory_order_relaxed);
-	g_primary_format = CapturedFormat{};

 	g_registered = 0;
 	g_streams_seen.store(0, std::memory_order_relaxed);
-	g_primary.store(nullptr, std::memory_order_release);
-	g_primary_block_align.store(0, std::memory_order_relaxed);
 	g_frames_captured.store(0, std::memory_order_relaxed);
-	for (auto& s : g_streams)
+	for (std::uint32_t i = 0; i < kMaxAudioStreams; ++i)
 	{
-		s.client.store(nullptr, std::memory_order_relaxed);
-		s.frames.store(0, std::memory_order_relaxed);
+		g_streams[i].client.store(nullptr, std::memory_order_relaxed);
+		g_streams[i].frames.store(0, std::memory_order_relaxed);
+		g_streams[i].block_align.store(0, std::memory_order_relaxed);
+		g_stream_formats[i] = CapturedFormat{};
+		g_rings[i].store(nullptr, std::memory_order_release);
 	}
 	g_client_formats.clear();
-	g_ring.store(nullptr, std::memory_order_release);
 	g_ipc = nullptr;
 }

--- a/hook/src/audio_hook.hpp
+++ b/hook/src/audio_hook.hpp
@@ -23,13 +23,15 @@ namespace coop::hook
 // anchor hook is in place; safe to call repeatedly (install-once internally).
 bool install_audio_hooks(IpcClient& ipc, AudioRingHeader* ring);

-// Attach/replace the producer ring after install (e.g. host created it late).
-void set_audio_ring(AudioRingHeader* ring);
+// Attach/replace the producer ring for stream `index` after install (e.g. host
+// created it late). Index 0 is the primary stream; the host creates one ring per
+// tracked stream (coop_audio_<pid>[_<index>]) and mixes them.
+void set_audio_ring(unsigned index, AudioRingHeader* ring);

-// Publish the registered primary stream's format to the attached ring if it
-// isn't published yet. Idempotent; call periodically so a ring the host attaches
-// (or re-initializes on a mirror re-toggle) gets the format even though the
-// stream was registered earlier. No-op if there's no ring / no primary yet.
+// Publish each registered stream's format to its attached ring if not published
+// yet. Idempotent; call periodically so a ring the host attaches (or re-initializes
+// on a mirror re-toggle) gets the format even though the stream was registered
+// earlier. No-op for rings that have no stream yet.
 void republish_audio_format();

 // Removes all installed render hooks (best effort; used on DLL detach).
--- a/hook/src/dllmain.cpp
+++ b/hook/src/dllmain.cpp
@@ -30,7 +30,7 @@ namespace

 coop::hook::IpcClient g_ipc;
 std::atomic<bool> g_running{true};
-coop::SharedMemory g_audio_shm; // the host's audio ring, opened when present
+coop::SharedMemory g_audio_shm[coop::kMaxAudioStreams]; // per-stream audio rings, opened when present
 coop::SharedMemory g_log_shm;                           // the host's log ring, opened when present

 DWORD WINAPI worker_thread(LPVOID)
@@ -168,31 +168,40 @@ DWORD WINAPI worker_thread(LPVOID)
 			coop::hook::logf("worker_thread: MKB hooks removed (host request)");
 		}

-		if (audio_installed && !audio_ring_open)
+		// Attach a ring per stream. The host creates up to kMaxAudioStreams rings
+		// (coop_audio_<pid>[_<index>]); we open each as it appears and (re)attach it so
+		// every stream is captured + silenced into its own ring for the host to mix.
+		if (audio_installed)
 		{
-			const std::wstring name = coop::audio_ring_name(GetCurrentProcessId());
-			if (!g_audio_shm.valid())
+			for (unsigned i = 0; i < coop::kMaxAudioStreams; ++i)
 			{
-				g_audio_shm.open(name, coop::audio_ring_total_size(coop::kAudioRingCapacity));
+				if (!g_audio_shm[i].valid())
+				{
+					g_audio_shm[i].open(coop::audio_ring_name(GetCurrentProcessId(), i),
+					                    coop::audio_ring_total_size(coop::kAudioRingCapacity));
 				}
-			if (g_audio_shm.valid())
+				if (g_audio_shm[i].valid())
 				{
-				auto* ring = g_audio_shm.as<coop::AudioRingHeader>();
+					auto* ring = g_audio_shm[i].as<coop::AudioRingHeader>();
 					if (coop::audio_ring_valid(*ring))
 					{
-					coop::hook::set_audio_ring(ring);
+						coop::hook::set_audio_ring(i, ring); // idempotent re-attach
+						if (i == 0 && !audio_ring_open)
+						{
 							audio_ring_open = true;
-					coop::hook::logf("worker_thread: audio ring opened");
+							coop::hook::logf("worker_thread: audio ring 0 opened");
+						}
 					}
 					else
 					{
-					g_audio_shm.reset(); // present but not our contract; retry
+						g_audio_shm[i].reset(); // present but not our contract; retry
 					}
 				}
 			}
-		// The primary stream is often registered before the ring is attached (or the
-		// host re-inits the ring on a mirror re-toggle, clearing its format); keep
-		// the format published so the host consumes the ring instead of falling back.
+		}
+		// A stream is often registered before its ring is attached (or the host re-inits
+		// a ring on a mirror re-toggle, clearing its format); keep formats published so
+		// the host consumes the rings instead of falling back to loopback.
 		if (audio_ring_open)
 		{
 			coop::hook::republish_audio_format();
--- a/host/src/audio/audio_loopback.cpp
+++ b/host/src/audio/audio_loopback.cpp
@@ -9,6 +9,7 @@
 #include <mmdeviceapi.h>
 #include <mmreg.h>

+#include "audio/audio_mix.hpp"
 #include "audio/process_loopback_capture.hpp"

 namespace coop
@@ -151,7 +152,10 @@ void AudioMirror::stop()
 		CloseHandle(stop_event_);
 		stop_event_ = nullptr;
 	}
-	audio_ring_shm_.reset();
+	for (auto& shm : audio_ring_shm_)
+	{
+		shm.reset();
+	}
 	running_.store(false, std::memory_order_release);
 	source_.store(Source::None, std::memory_order_relaxed);
 	buffered_ms_.store(0, std::memory_order_relaxed);
@@ -191,15 +195,24 @@ void AudioMirror::thread_main(DWORD pid)
 	// capture. If the hook is present it publishes a format within ~1 s and we
 	// consume the ring (no echo); otherwise we fall back to process loopback.
 	bool handled = false;
-	if (audio_ring_shm_.create(audio_ring_name(pid), audio_ring_total_size(kAudioRingCapacity)))
+	AudioRingHeader* rings[kMaxAudioStreams] = {};
+	bool created_primary = false;
+	for (unsigned i = 0; i < kMaxAudioStreams; ++i)
+	{
+		if (audio_ring_shm_[i].create(audio_ring_name(pid, i), audio_ring_total_size(kAudioRingCapacity)))
+		{
+			rings[i] = audio_ring_shm_[i].as<AudioRingHeader>();
+			audio_ring_init(*rings[i], kAudioRingCapacity);
+			rings[i]->capture_enabled.store(1, std::memory_order_release);
+			created_primary = created_primary || (i == 0);
+		}
+	}
+	if (created_primary)
 	{
-		auto* ring = audio_ring_shm_.as<AudioRingHeader>();
-		audio_ring_init(*ring, kAudioRingCapacity);
-		ring->capture_enabled.store(1, std::memory_order_release);
 		set_status("Waiting for render-hook…");
-		if (wait_for_format(ring, 1000))
+		if (wait_for_format(rings[0], 1000))
 		{
-			handled = run_hooked(ring);
+			handled = run_hooked(rings);
 		}
 	}

@@ -208,7 +221,10 @@ void AudioMirror::thread_main(DWORD pid)
 		run_loopback(pid);
 	}

-	audio_ring_shm_.reset();
+	for (auto& shm : audio_ring_shm_)
+	{
+		shm.reset();
+	}

 	if (running_.load(std::memory_order_acquire))
 	{
@@ -226,17 +242,28 @@ void AudioMirror::thread_main(DWORD pid)
 // Consume the render-hook's shared ring and re-render the game's frames. The
 // game is silenced locally by the hook, so the operator hears no echo. Returns
 // true if it ran to a clean stop; false on setup failure (caller falls back).
-bool AudioMirror::run_hooked(AudioRingHeader* ring)
+bool AudioMirror::run_hooked(AudioRingHeader* const* rings)
 {
+	AudioRingHeader* primary = rings[0];
+	auto disable_all = [&] {
+		for (unsigned i = 0; i < kMaxAudioStreams; ++i)
+		{
+			if (rings[i] != nullptr)
+			{
+				rings[i]->capture_enabled.store(0, std::memory_order_release);
+			}
+		}
+	};
 	auto fail_to_loopback = [&] {
-		ring->capture_enabled.store(0, std::memory_order_release); // let the game play locally again
+		disable_all(); // let the game play locally again
 		return false;
 	};

-	const unsigned rate = ring->sample_rate;
-	const unsigned channels = ring->channels;
-	const unsigned bits = ring->bits;
-	const unsigned block_align = ring->block_align ? ring->block_align : channels * (bits / 8);
+	const unsigned rate = primary->sample_rate;
+	const unsigned channels = primary->channels;
+	const unsigned bits = primary->bits;
+	const unsigned tag = primary->format_tag;
+	const unsigned block_align = primary->block_align ? primary->block_align : channels * (bits / 8);
 	if (rate == 0 || channels == 0 || block_align == 0)
 	{
 		return fail_to_loopback();
@@ -267,12 +294,12 @@ bool AudioMirror::run_hooked(AudioRingHeader* ring)
 			wfx.dwChannelMask = (channels >= 32) ? 0xFFFFFFFFu : ((1u << channels) - 1u);
 			break;
 		}
-		wfx.SubFormat = (ring->format_tag == WAVE_FORMAT_IEEE_FLOAT) ? KSDATAFORMAT_SUBTYPE_IEEE_FLOAT
-		                                                             : KSDATAFORMAT_SUBTYPE_PCM;
+		wfx.SubFormat =
+		    (tag == WAVE_FORMAT_IEEE_FLOAT) ? KSDATAFORMAT_SUBTYPE_IEEE_FLOAT : KSDATAFORMAT_SUBTYPE_PCM;
 	}
 	else
 	{
-		wfx.Format.wFormatTag = static_cast<WORD>(ring->format_tag ? ring->format_tag : WAVE_FORMAT_PCM);
+		wfx.Format.wFormatTag = static_cast<WORD>(tag ? tag : WAVE_FORMAT_PCM);
 		wfx.Format.cbSize = 0;
 	}
 	auto* fmt = reinterpret_cast<WAVEFORMATEX*>(&wfx);
@@ -355,6 +382,12 @@ bool AudioMirror::run_hooked(AudioRingHeader* ring)
 		const size_t prime_bytes = frame_bytes * (rate * 30 / 1000); // ~30 ms before feeding
 		bool primed = false;

+		// Mixing scratch (only used when >1 same-format stream is active): a per-stream
+		// temp buffer and a float accumulator sized to the render buffer.
+		const bool mixer_ok = mix_format_supported(tag, bits);
+		std::vector<BYTE> temp(static_cast<size_t>(render_frames) * frame_bytes);
+		std::vector<float> acc(static_cast<size_t>(render_frames) * channels);
+
 		if (FAILED(hr = render_client->Start()))
 		{
 			fail("Render Start", hr);
@@ -380,7 +413,8 @@ bool AudioMirror::run_hooked(AudioRingHeader* ring)
 				continue;
 			}
 			const UINT32 avail = render_frames - padding;
-			const std::uint32_t ring_bytes = audio_ring_available(*ring);
+			// The primary ring is the master clock for priming + how much to write.
+			const std::uint32_t ring_bytes = audio_ring_available(*primary);
 			buffered_ms_.store(static_cast<unsigned>(ring_bytes / frame_bytes * 1000 / rate),
 			                   std::memory_order_relaxed);
 			if (!primed && ring_bytes >= prime_bytes)
@@ -393,10 +427,47 @@ bool AudioMirror::run_hooked(AudioRingHeader* ring)
 				const UINT32 to_write = std::min(avail, have);
 				if (to_write > 0)
 				{
+					// Active streams = same format as primary (so they can be summed).
+					// Streams with a different format are still silenced by the hook (no
+					// echo) but can't be mixed here without resampling -> skipped.
+					unsigned active[kMaxAudioStreams];
+					unsigned n_active = 0;
+					for (unsigned i = 0; i < kMaxAudioStreams; ++i)
+					{
+						AudioRingHeader* r = rings[i];
+						if (r == nullptr)
+						{
+							continue;
+						}
+						if (i == 0 || (audio_ring_format_ready(*r) && r->sample_rate == rate &&
+						               r->channels == channels && r->bits == bits && r->format_tag == tag))
+						{
+							active[n_active++] = i;
+						}
+					}
+
 					BYTE* dst = nullptr;
 					if (SUCCEEDED(render->GetBuffer(to_write, &dst)))
 					{
-						audio_ring_pop(*ring, dst, to_write * static_cast<std::uint32_t>(frame_bytes));
+						const std::uint32_t want_bytes = to_write * static_cast<std::uint32_t>(frame_bytes);
+						if (n_active <= 1 || !mixer_ok)
+						{
+							// Single stream (the common case) or an unmixable format:
+							// passthrough the primary, byte-for-byte (no mixer overhead).
+							audio_ring_pop(*primary, dst, want_bytes);
+						}
+						else
+						{
+							const std::uint32_t samples = to_write * channels;
+							std::fill(acc.begin(), acc.begin() + samples, 0.0f);
+							for (unsigned k = 0; k < n_active; ++k)
+							{
+								std::memset(temp.data(), 0, want_bytes); // zero-fill short reads
+								audio_ring_pop(*rings[active[k]], temp.data(), want_bytes);
+								mix_add(acc.data(), temp.data(), samples, tag, bits);
+							}
+							mix_store(dst, acc.data(), samples, tag, bits);
+						}
 						render->ReleaseBuffer(to_write, 0);
 					}
 				}
@@ -410,7 +481,7 @@ bool AudioMirror::run_hooked(AudioRingHeader* ring)
 		render_client->Stop();
 	} while (false);

-	ring->capture_enabled.store(0, std::memory_order_release); // game audible again on stop
+	disable_all(); // game audible again on stop

 	if (render)
 	{
--- a/host/src/audio/audio_loopback.hpp
+++ b/host/src/audio/audio_loopback.hpp
@@ -18,6 +18,7 @@
 #include <windows.h>

 #include "coop/audio_ring.hpp"
+#include "coop/protocol.hpp" // kMaxAudioStreams
 #include "coop/shared_memory.hpp"

 namespace coop
@@ -97,8 +98,9 @@ public:
 private:
 	void thread_main(DWORD pid);
 	// Returns true if it owned the session to a clean stop; false if setup failed
-	// and the caller should fall back to the loopback path.
-	bool run_hooked(AudioRingHeader* ring);
+	// and the caller should fall back to the loopback path. `rings[0]` is the primary
+	// stream; additional non-null rings are mixed in.
+	bool run_hooked(AudioRingHeader* const* rings);
 	void run_loopback(DWORD pid);
 	bool wait_for_format(AudioRingHeader* ring, DWORD timeout_ms);
 	bool stop_requested() const;
@@ -108,7 +110,7 @@ private:
 	HANDLE stop_event_ = nullptr;
 	DWORD pid_ = 0;

-	SharedMemory audio_ring_shm_; // host-created shared ring (named coop_audio_<pid>)
+	SharedMemory audio_ring_shm_[kMaxAudioStreams]; // per-stream rings (coop_audio_<pid>[_<i>])

 	std::atomic<bool> running_{false};
 	std::atomic<Source> source_{Source::None};
--- a/host/src/audio/audio_mix.hpp
+++ b/host/src/audio/audio_mix.hpp
@@ -0,0 +1,77 @@
+// Pure helpers for mixing several same-format audio streams (float32 or int16)
+// into one render buffer, with a soft clip on the sum. Header-only so the host
+// mixer and a unit test share the exact same math.
+#pragma once
+
+#include <cmath>
+#include <cstdint>
+
+namespace coop
+{
+
+// WAVE_FORMAT_* values used here (kept local to avoid an mmreg.h dependency).
+inline constexpr std::uint32_t kWaveFormatPcm = 1;
+inline constexpr std::uint32_t kWaveFormatFloat = 3;
+
+// The mixer decodes/encodes these formats; anything else falls back to a plain
+// single-stream passthrough on the host (no mixing).
+inline bool mix_format_supported(std::uint32_t format_tag, std::uint32_t bits)
+{
+	return (format_tag == kWaveFormatFloat && bits == 32) || (format_tag == kWaveFormatPcm && bits == 16);
+}
+
+// Smoothly limit the summed signal to [-1, 1]; ~identity for small inputs.
+inline float soft_clip(float x)
+{
+	return std::tanh(x);
+}
+
+// Decode `samples` interleaved samples from `src` (format_tag/bits) and add them
+// into the float accumulator `acc`. No-op for unsupported formats.
+inline void mix_add(float* acc, const std::uint8_t* src, std::uint32_t samples, std::uint32_t format_tag,
+                    std::uint32_t bits)
+{
+	if (format_tag == kWaveFormatFloat && bits == 32)
+	{
+		const auto* f = reinterpret_cast<const float*>(src);
+		for (std::uint32_t i = 0; i < samples; ++i)
+		{
+			acc[i] += f[i];
+		}
+	}
+	else if (format_tag == kWaveFormatPcm && bits == 16)
+	{
+		const auto* s = reinterpret_cast<const std::int16_t*>(src);
+		for (std::uint32_t i = 0; i < samples; ++i)
+		{
+			acc[i] += static_cast<float>(s[i]) / 32768.0f;
+		}
+	}
+}
+
+// Soft-clip `acc` and encode `samples` interleaved samples into `dst` in the given
+// format. No-op for unsupported formats.
+inline void mix_store(std::uint8_t* dst, const float* acc, std::uint32_t samples, std::uint32_t format_tag,
+                      std::uint32_t bits)
+{
+	if (format_tag == kWaveFormatFloat && bits == 32)
+	{
+		auto* f = reinterpret_cast<float*>(dst);
+		for (std::uint32_t i = 0; i < samples; ++i)
+		{
+			f[i] = soft_clip(acc[i]);
+		}
+	}
+	else if (format_tag == kWaveFormatPcm && bits == 16)
+	{
+		auto* s = reinterpret_cast<std::int16_t*>(dst);
+		for (std::uint32_t i = 0; i < samples; ++i)
+		{
+			int v = static_cast<int>(soft_clip(acc[i]) * 32767.0f);
+			v = v > 32767 ? 32767 : (v < -32768 ? -32768 : v);
+			s[i] = static_cast<std::int16_t>(v);
+		}
+	}
+}
+
+} // namespace coop
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -25,6 +25,11 @@ add_executable(mkb_ring_test mkb_ring_test.cpp)
 target_link_libraries(mkb_ring_test PRIVATE coop_common)
 add_test(NAME mkb_ring_test COMMAND mkb_ring_test)

+# Unit test for the audio mixer math (decode/sum/soft-clip/encode). Header-only.
+add_executable(audio_mix_test audio_mix_test.cpp)
+target_include_directories(audio_mix_test PRIVATE ${CMAKE_SOURCE_DIR}/host/src)
+add_test(NAME audio_mix_test COMMAND audio_mix_test)
+
 # Unit test for the host->game mouse coordinate mapping (letterbox inverse +
 # decorated-window client offset). Header-only, no device.
 add_executable(mkb_map_test mkb_map_test.cpp)
@@ -148,6 +153,7 @@ coop_output_subdir(tests
 	audio_ring_test
 	mkb_ring_test
 	mkb_map_test
+	audio_mix_test
 	audio_loopback_test
 	audio_hook_test
 	srgb_format_test
--- a/tests/audio_mix_test.cpp
+++ b/tests/audio_mix_test.cpp
@@ -0,0 +1,89 @@
+// Unit test for the audio mixer math (decode/sum/soft-clip/encode, float32 + int16).
+#include <cmath>
+#include <cstdint>
+#include <cstdio>
+#include <cstring>
+#include <vector>
+
+#include "audio/audio_mix.hpp"
+
+using namespace coop;
+
+namespace
+{
+int g_failures = 0;
+void check(bool ok, const char* what)
+{
+	if (!ok)
+	{
+		std::printf("FAIL: %s\n", what);
+		++g_failures;
+	}
+}
+bool near_f(float a, float b)
+{
+	return std::fabs(a - b) < 1e-4f;
+}
+} // namespace
+
+int main()
+{
+	check(mix_format_supported(kWaveFormatFloat, 32), "float32 supported");
+	check(mix_format_supported(kWaveFormatPcm, 16), "int16 supported");
+	check(!mix_format_supported(kWaveFormatPcm, 24), "24-bit not supported");
+
+	// soft_clip is ~identity for small inputs and bounded for large ones.
+	check(near_f(soft_clip(0.0f), 0.0f), "soft_clip(0)=0");
+	check(soft_clip(10.0f) <= 1.0f && soft_clip(10.0f) > 0.99f, "soft_clip bounds large +");
+	check(soft_clip(-10.0f) >= -1.0f && soft_clip(-10.0f) < -0.99f, "soft_clip bounds large -");
+
+	// --- float32: two streams sum, small values pass ~unchanged ---
+	{
+		const float a[4] = {0.1f, -0.2f, 0.3f, -0.05f};
+		const float b[4] = {0.2f, 0.1f, -0.1f, 0.05f};
+		float acc[4] = {};
+		mix_add(acc, reinterpret_cast<const std::uint8_t*>(a), 4, kWaveFormatFloat, 32);
+		mix_add(acc, reinterpret_cast<const std::uint8_t*>(b), 4, kWaveFormatFloat, 32);
+		float out[4] = {};
+		mix_store(reinterpret_cast<std::uint8_t*>(out), acc, 4, kWaveFormatFloat, 32);
+		// Sum then tanh; small sums are ~unchanged.
+		for (int i = 0; i < 4; ++i)
+		{
+			check(near_f(out[i], std::tanh(a[i] + b[i])), "float32 mix == tanh(sum)");
+		}
+	}
+
+	// --- float32: summing many loud streams stays within [-1, 1] (soft clip) ---
+	{
+		float acc[2] = {};
+		const float loud[2] = {0.9f, -0.9f};
+		for (int s = 0; s < 5; ++s)
+		{
+			mix_add(acc, reinterpret_cast<const std::uint8_t*>(loud), 2, kWaveFormatFloat, 32);
+		}
+		float out[2] = {};
+		mix_store(reinterpret_cast<std::uint8_t*>(out), acc, 2, kWaveFormatFloat, 32);
+		check(out[0] <= 1.0f && out[0] > 0.99f, "loud sum soft-clipped near +1");
+		check(out[1] >= -1.0f && out[1] < -0.99f, "loud sum soft-clipped near -1");
+	}
+
+	// --- int16: decode/encode round-trip of a single quiet stream ---
+	{
+		const std::int16_t a[2] = {1000, -2000};
+		float acc[2] = {};
+		mix_add(acc, reinterpret_cast<const std::uint8_t*>(a), 2, kWaveFormatPcm, 16);
+		check(near_f(acc[0], 1000.0f / 32768.0f) && near_f(acc[1], -2000.0f / 32768.0f), "int16 decode");
+		std::int16_t out[2] = {};
+		mix_store(reinterpret_cast<std::uint8_t*>(out), acc, 2, kWaveFormatPcm, 16);
+		// tanh of a tiny value ~ the value, so re-encoding is within a couple of LSB.
+		check(std::abs(out[0] - 1000) <= 3 && std::abs(out[1] - (-2000)) <= 3, "int16 round-trip");
+	}
+
+	if (g_failures == 0)
+	{
+		std::printf("PASS: audio_mix_test\n");
+		return 0;
+	}
+	std::printf("FAIL: %d checks\n", g_failures);
+	return 1;
+}