diff --git a/nx/include/switch/services/hwopus.h b/nx/include/switch/services/hwopus.h
index 1d6c54c3..4e6abfa3 100644
--- a/nx/include/switch/services/hwopus.h
+++ b/nx/include/switch/services/hwopus.h
@@ -12,6 +12,7 @@
 typedef struct {
     Service s;
     TransferMemory tmem;
+    bool multistream;
 } HwopusDecoder;
 
 ///< This structure is the start of opusin for \ref hwopusDecodeInterleaved, with the actual opus packet following this.
diff --git a/nx/source/services/hwopus.c b/nx/source/services/hwopus.c
index 1b55ad94..88cb9098 100644
--- a/nx/source/services/hwopus.c
+++ b/nx/source/services/hwopus.c
@@ -10,6 +10,8 @@
 static Result _hwopusInitialize(Service* srv, Service* out_srv, TransferMemory *tmem, s32 SampleRate, s32 ChannelCount);
 static Result _hwopusGetWorkBufferSize(Service* srv, u32 *size, s32 SampleRate, s32 ChannelCount);
 
+static Result _hwopusDecodeInterleavedWithPerfOld(HwopusDecoder* decoder, s32 *DecodedDataSize, s32 *DecodedSampleCount, u64 *perf, const void* opusin, size_t opusin_size, s16 *pcmbuf, size_t pcmbuf_size);
+
 Result hwopusDecoderInitialize(HwopusDecoder* decoder, s32 SampleRate, s32 ChannelCount) {
     Result rc=0;
     u32 size=0;
@@ -17,6 +19,8 @@ Result hwopusDecoderInitialize(HwopusDecoder* decoder, s32 SampleRate, s32 Chann
     if (serviceIsActive(&decoder->s))
         return 0;
 
+    decoder->multistream = false;
+
     Service hwopusMgrSrv;
     rc = smGetService(&hwopusMgrSrv, "hwopus");
     if (R_SUCCEEDED(rc)) {
@@ -121,6 +125,8 @@ static Result _hwopusGetWorkBufferSize(Service* srv, u32 *size, s32 SampleRate,
 }
 
 Result hwopusDecodeInterleaved(HwopusDecoder* decoder, s32 *DecodedDataSize, s32 *DecodedSampleCount, const void* opusin, size_t opusin_size, s16 *pcmbuf, size_t pcmbuf_size) {
+    if (kernelAbove400()) return _hwopusDecodeInterleavedWithPerfOld(decoder, DecodedDataSize, DecodedSampleCount, NULL, opusin, opusin_size, pcmbuf, pcmbuf_size);
+
     IpcCommand c;
     ipcInitialize(&c);
 
@@ -135,7 +141,7 @@ Result hwopusDecodeInterleaved(HwopusDecoder* decoder, s32 *DecodedDataSize, s32
     raw = serviceIpcPrepareHeader(&decoder->s, &c, sizeof(*raw));
 
     raw->magic = SFCI_MAGIC;
-    raw->cmd_id = 0;
+    raw->cmd_id = decoder->multistream==0 ? 0 : 2;
 
     Result rc = serviceIpcDispatch(&decoder->s);
 
@@ -159,3 +165,45 @@ Result hwopusDecodeInterleaved(HwopusDecoder* decoder, s32 *DecodedDataSize, s32
 
     return rc;
 }
+
+static Result _hwopusDecodeInterleavedWithPerfOld(HwopusDecoder* decoder, s32 *DecodedDataSize, s32 *DecodedSampleCount, u64 *perf, const void* opusin, size_t opusin_size, s16 *pcmbuf, size_t pcmbuf_size) {
+    IpcCommand c;
+    ipcInitialize(&c);
+
+    ipcAddSendBuffer(&c, opusin, opusin_size, BufferType_Normal);
+    ipcAddRecvBuffer(&c, pcmbuf, pcmbuf_size, BufferType_Type1);
+
+    struct {
+        u64 magic;
+        u64 cmd_id;
+    } *raw;
+
+    raw = serviceIpcPrepareHeader(&decoder->s, &c, sizeof(*raw));
+
+    raw->magic = SFCI_MAGIC;
+    raw->cmd_id = decoder->multistream==0 ? 4 : 5;
+
+    Result rc = serviceIpcDispatch(&decoder->s);
+
+    if (R_SUCCEEDED(rc)) {
+        IpcParsedCommand r;
+        struct {
+            u64 magic;
+            u64 result;
+            s32 DecodedDataSize;
+            s32 DecodedSampleCount;
+            u64 perf;
+        } *resp;
+
+        serviceIpcParse(&decoder->s, &r, sizeof(*resp));
+        resp = r.Raw;
+
+        rc = resp->result;
+
+        if (R_SUCCEEDED(rc) && DecodedSampleCount) *DecodedSampleCount = resp->DecodedSampleCount;
+        if (R_SUCCEEDED(rc) && DecodedDataSize) *DecodedDataSize = resp->DecodedDataSize;
+        if (R_SUCCEEDED(rc) && perf) *perf = resp->perf;
+    }
+
+    return rc;
+}