diff --git a/include/lol/private/audio/stream.h b/include/lol/private/audio/stream.h
index 1b1d8666..bfe79f25 100644
--- a/include/lol/private/audio/stream.h
+++ b/include/lol/private/audio/stream.h
@@ -35,9 +35,7 @@ public:
     static inline TO convert(FROM x)
     {
         constexpr auto from_fp = std::is_floating_point_v<FROM>;
-        constexpr auto from_signed = !from_fp && std::is_signed_v<FROM>;
         constexpr auto to_fp = std::is_floating_point_v<TO>;
-        constexpr auto to_signed = !to_fp && std::is_signed_v<TO>;
 
         if constexpr (std::is_same_v<FROM, TO> || (from_fp && to_fp))
         {
@@ -47,7 +45,7 @@ public:
         else if constexpr (from_fp)
         {
             // From floating point to integer:
-            //  - renormalise to 0…1
+            //  - change range from -1…1 to 0…1
             //  - multiply by the size of the integer range
             //  - add min, round down, and clamp to min…max
             FROM constexpr min(std::numeric_limits<TO>::min());
@@ -57,14 +55,38 @@ public:
         }
         else if constexpr (to_fp)
         {
+            // From integer to floating point:
+            //  - compute (x - min) / (max - min)
+            //  - change range from 0…1 to -1…1
             TO constexpr min(std::numeric_limits<FROM>::min());
             TO constexpr max(std::numeric_limits<FROM>::max());
             return (TO(x) - min) * 2 / (max - min) - 1;
         }
         else
         {
-            // FIXME: this is better than nothing but we need a better implementation
-            return convert<double, TO>(convert<FROM, double>(x));
+            using UFROM = std::make_unsigned_t<FROM>;
+            using UTO = std::make_unsigned_t<TO>;
+
+            if constexpr (sizeof(FROM) > sizeof(TO))
+            {
+                // From a larger integer type to a smaller integer type:
+                //  - convert to unsigned
+                //  - shift right
+                //  - convert back to signed if necessary
+                UFROM constexpr m = UFROM(1) << (8 * (sizeof(FROM) - sizeof(TO)));
+                UFROM tmp = UFROM(UFROM(x) - UFROM(std::numeric_limits<FROM>::min())) / m;
+                return TO(UTO(tmp) + UTO(std::numeric_limits<TO>::min()));
+            }
+            else
+            {
+                // From a smaller integer type to a larger integer type:
+                //  - convert to unsigned
+                //  - multiply by a magic constant such as 0x01010101 to propagate bytes
+                //  - convert back to signed if necessary
+                UTO constexpr m = std::numeric_limits<UTO>::max() / std::numeric_limits<UFROM>::max();
+                UTO tmp = UFROM(UFROM(x) - UFROM(std::numeric_limits<FROM>::min())) * m;
+                return TO(UTO(tmp) + UTO(std::numeric_limits<TO>::min()));
+            }
         }
     }
 };
diff --git a/t/audio.cpp b/t/audio.cpp
index 5dc32714..5abd31cd 100644
--- a/t/audio.cpp
+++ b/t/audio.cpp
@@ -27,81 +27,123 @@ TEST_CASE("sample conversion between floating point types")
 TEST_CASE("sample conversion from float to 8-bit")
 {
     auto cv1 = lol::audio::sample::convert<float, int8_t>;
-    CHECK(cv1(-1.5f) == -128);
-    CHECK(cv1(-1.0f) == -128);
-    CHECK(cv1(-0.5f) == -64);
-    CHECK(cv1( 0.0f) ==  0);
-    CHECK(cv1( 0.5f) ==  64);
-    CHECK(cv1( 1.0f) ==  127);
-    CHECK(cv1( 1.5f) ==  127);
+    CHECK(cv1(-1.5f) == -0x80);
+    CHECK(cv1(-1.0f) == -0x80);
+    CHECK(cv1(-0.5f) == -0x40);
+    CHECK(cv1( 0.0f) ==  0x00);
+    CHECK(cv1( 0.5f) ==  0x40);
+    CHECK(cv1( 1.0f) ==  0x7f);
+    CHECK(cv1( 1.5f) ==  0x7f);
 
     auto cv2 = lol::audio::sample::convert<float, uint8_t>;
-    CHECK(cv2(-1.5f) == 0);
-    CHECK(cv2(-1.0f) == 0);
-    CHECK(cv2(-0.5f) == 64);
-    CHECK(cv2( 0.0f) == 128);
-    CHECK(cv2( 0.5f) == 192);
-    CHECK(cv2( 1.0f) == 255);
-    CHECK(cv2( 1.5f) == 255);
+    CHECK(cv2(-1.5f) == 0x00);
+    CHECK(cv2(-1.0f) == 0x00);
+    CHECK(cv2(-0.5f) == 0x40);
+    CHECK(cv2( 0.0f) == 0x80);
+    CHECK(cv2( 0.5f) == 0xc0);
+    CHECK(cv2( 1.0f) == 0xff);
+    CHECK(cv2( 1.5f) == 0xff);
 }
 
 TEST_CASE("sample conversion from 8-bit to float")
 {
     auto cv1 = lol::audio::sample::convert<int8_t, float>;
-    CHECK(cv1(-128) == -1.0f);
-    CHECK(cv1( 127) ==  1.0f);
+    CHECK(cv1(-0x80) == -1.0f);
+    CHECK(cv1( 0x7f) ==  1.0f);
 
     auto cv2 = lol::audio::sample::convert<uint8_t, float>;
-    CHECK(cv2(  0) == -1.0f);
-    CHECK(cv2(255) ==  1.0f);
+    CHECK(cv2(0x00) == -1.0f);
+    CHECK(cv2(0xff) ==  1.0f);
 }
 
 TEST_CASE("sample conversion from float to 16-bit")
 {
     auto cv1 = lol::audio::sample::convert<float, int16_t>;
-    CHECK(cv1(-1.5f) == -32768);
-    CHECK(cv1(-1.0f) == -32768);
-    CHECK(cv1(-0.5f) == -16384);
-    CHECK(cv1( 0.0f) ==  0);
-    CHECK(cv1( 0.5f) ==  16384);
-    CHECK(cv1( 1.0f) ==  32767);
-    CHECK(cv1( 1.5f) ==  32767);
+    CHECK(cv1(-1.5f) == -0x8000);
+    CHECK(cv1(-1.0f) == -0x8000);
+    CHECK(cv1(-0.5f) == -0x4000);
+    CHECK(cv1( 0.0f) ==  0x0000);
+    CHECK(cv1( 0.5f) ==  0x4000);
+    CHECK(cv1( 1.0f) ==  0x7fff);
+    CHECK(cv1( 1.5f) ==  0x7fff);
 
     auto cv2 = lol::audio::sample::convert<float, uint16_t>;
-    CHECK(cv2(-1.5f) == 0);
-    CHECK(cv2(-1.0f) == 0);
-    CHECK(cv2(-0.5f) == 16384);
-    CHECK(cv2( 0.0f) == 32768);
-    CHECK(cv2( 0.5f) == 49152);
-    CHECK(cv2( 1.0f) == 65535);
-    CHECK(cv2( 1.5f) == 65535);
+    CHECK(cv2(-1.5f) == 0x0000);
+    CHECK(cv2(-1.0f) == 0x0000);
+    CHECK(cv2(-0.5f) == 0x4000);
+    CHECK(cv2( 0.0f) == 0x8000);
+    CHECK(cv2( 0.5f) == 0xc000);
+    CHECK(cv2( 1.0f) == 0xffff);
+    CHECK(cv2( 1.5f) == 0xffff);
 }
 
 TEST_CASE("sample conversion from 16-bit to float")
 {
     auto cv1 = lol::audio::sample::convert<int16_t, float>;
-    CHECK(cv1(-32768) == -1.0f);
-    CHECK(cv1( 32767) ==  1.0f);
+    CHECK(cv1(-0x8000) == -1.0f);
+    CHECK(cv1( 0x7fff) ==  1.0f);
 
     auto cv2 = lol::audio::sample::convert<uint16_t, float>;
-    CHECK(cv2(    0) == -1.0f);
-    CHECK(cv2(65535) ==  1.0f);
+    CHECK(cv2(0x0000) == -1.0f);
+    CHECK(cv2(0xffff) ==  1.0f);
+}
+
+TEST_CASE("sample conversion between signed and unsigned 8-bit")
+{
+    auto cv1 = lol::audio::sample::convert<int8_t, uint8_t>;
+    CHECK(cv1(-0x80) == 0x00);
+    CHECK(cv1(-0x40) == 0x40);
+    CHECK(cv1(-0x01) == 0x7f);
+    CHECK(cv1( 0x00) == 0x80);
+    CHECK(cv1( 0x3f) == 0xbf);
+    CHECK(cv1( 0x7f) == 0xff);
+
+    auto cv2 = lol::audio::sample::convert<uint8_t, int8_t>;
+    CHECK(cv2(0x00) == -0x80);
+    CHECK(cv2(0x40) == -0x40);
+    CHECK(cv2(0x7f) == -0x01);
+    CHECK(cv2(0x80) ==  0x00);
+    CHECK(cv2(0xbf) ==  0x3f);
+    CHECK(cv2(0xff) ==  0x7f);
+}
+
+TEST_CASE("sample conversion from 16-bit to 8-bit")
+{
+    auto cv1 = lol::audio::sample::convert<int16_t, int8_t>;
+    CHECK(cv1(-0x8000) == -0x80);
+    CHECK(cv1(-0x4000) == -0x40);
+    CHECK(cv1(-0x0080) == -0x01);
+    CHECK(cv1(-0x0001) == -0x01);
+    CHECK(cv1( 0x0000) ==  0x00);
+    CHECK(cv1( 0x00ff) ==  0x00);
+    CHECK(cv1( 0x3fff) ==  0x3f);
+    CHECK(cv1( 0x7fff) ==  0x7f);
+
+    auto cv2 = lol::audio::sample::convert<uint16_t, int8_t>;
+    CHECK(cv2(0x0000) == -0x80);
+    CHECK(cv2(0x4000) == -0x40);
+    CHECK(cv2(0x7f80) == -0x01);
+    CHECK(cv2(0x7fff) == -0x01);
+    CHECK(cv2(0x8000) ==  0x00);
+    CHECK(cv2(0x80ff) ==  0x00);
+    CHECK(cv2(0xbfff) ==  0x3f);
+    CHECK(cv2(0xffff) ==  0x7f);
 }
 
 TEST_CASE("round-trip conversion from 8-bit to 8-bit")
 {
     auto cv1 = lol::audio::sample::convert<int8_t, float>;
     auto cv2 = lol::audio::sample::convert<float, int8_t>;
-    CHECK(cv2(cv1(-128)) == -128);
-    CHECK(cv2(cv1(-127)) == -127);
-    CHECK(cv2(cv1( -64)) ==  -64);
-    CHECK(cv2(cv1( -32)) ==  -32);
-    CHECK(cv2(cv1(  -2)) ==   -2);
-    CHECK(cv2(cv1(  -1)) ==   -1);
-    CHECK(cv2(cv1(   0)) ==    0);
-    CHECK(cv2(cv1(   1)) ==    1);
-    CHECK(cv2(cv1(   2)) ==    2);
-    CHECK(cv2(cv1(  32)) ==   32);
-    CHECK(cv2(cv1(  64)) ==   64);
-    CHECK(cv2(cv1( 127)) ==  127);
+    CHECK(cv2(cv1(-0x80)) == -0x80);
+    CHECK(cv2(cv1(-0x7f)) == -0x7f);
+    CHECK(cv2(cv1(-0x40)) == -0x40);
+    CHECK(cv2(cv1(-0x20)) == -0x20);
+    CHECK(cv2(cv1(-0x02)) == -0x02);
+    CHECK(cv2(cv1(-0x01)) == -0x01);
+    CHECK(cv2(cv1( 0x00)) ==  0x00);
+    CHECK(cv2(cv1( 0x01)) ==  0x01);
+    CHECK(cv2(cv1( 0x02)) ==  0x02);
+    CHECK(cv2(cv1( 0x20)) ==  0x20);
+    CHECK(cv2(cv1( 0x40)) ==  0x40);
+    CHECK(cv2(cv1( 0x7f)) ==  0x7f);
 }