diff --git a/CHANGELOG.md b/CHANGELOG.md
index 5cb66e2d..69f8f170 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,7 @@
   yielding samples when the underlying signal gets exhausted. This is a breaking
   change. The return type of the `IntoInterleavedSamples#next_sample` method was
   modified.
+- Improved float-to-integer conversions to use proper rounding instead of truncation.
 
 ---
 
diff --git a/dasp_sample/src/conv.rs b/dasp_sample/src/conv.rs
index f1ed3f90..57c9bbf8 100644
--- a/dasp_sample/src/conv.rs
+++ b/dasp_sample/src/conv.rs
@@ -126,7 +126,9 @@ macro_rules! conversion_fns {
 macro_rules! conversions {
     ($T:ident, $mod_name:ident { $($rest:tt)* }) => {
         pub mod $mod_name {
-            use $crate::types::{I24, U24, I48, U48};
+            #[allow(unused_imports)]
+            use $crate::ops;
+            use $crate::{types::{I24, U24, I48, U48}};
             conversion_fns!($T, $($rest)*);
         }
     };
@@ -531,12 +533,12 @@ conversions!(u64, u64 {
 // The following conversions assume `-1.0 <= s < 1.0` (note that +1.0 is excluded) and will
 // overflow otherwise.
 conversions!(f32, f32 {
-    s to_i8 { (s * 128.0) as i8 }
-    s to_i16 { (s * 32_768.0) as i16 }
-    s to_i24 { I24::new_unchecked((s * 8_388_608.0) as i32) }
-    s to_i32 { (s * 2_147_483_648.0) as i32 }
-    s to_i48 { I48::new_unchecked((s * 140_737_488_355_328.0) as i64) }
-    s to_i64 { (s * 9_223_372_036_854_775_808.0) as i64 }
+    s to_i8 { ops::f32::round(s * 128.0) as i8 }
+    s to_i16 { ops::f32::round(s * 32_768.0) as i16 }
+    s to_i24 { I24::new_unchecked(ops::f32::round(s * 8_388_608.0) as i32) }
+    s to_i32 { ops::f32::round(s * 2_147_483_648.0) as i32 }
+    s to_i48 { I48::new_unchecked(ops::f32::round(s * 140_737_488_355_328.0) as i64) }
+    s to_i64 { ops::f32::round(s * 9_223_372_036_854_775_808.0) as i64 }
     s to_u8 { super::i8::to_u8(to_i8(s)) }
     s to_u16 { super::i16::to_u16(to_i16(s)) }
     s to_u24 { super::i24::to_u24(to_i24(s)) }
@@ -549,12 +551,12 @@ conversions!(f32, f32 {
 // The following conversions assume `-1.0 <= s < 1.0` (note that +1.0 is excluded) and will
 // overflow otherwise.
 conversions!(f64, f64 {
-    s to_i8 { (s * 128.0) as i8 }
-    s to_i16 { (s * 32_768.0) as i16 }
-    s to_i24 { I24::new_unchecked((s * 8_388_608.0) as i32) }
-    s to_i32 { (s * 2_147_483_648.0) as i32 }
-    s to_i48 { I48::new_unchecked((s * 140_737_488_355_328.0) as i64) }
-    s to_i64 { (s * 9_223_372_036_854_775_808.0) as i64 }
+    s to_i8 { ops::f64::round(s * 128.0) as i8 }
+    s to_i16 { ops::f64::round(s * 32_768.0) as i16 }
+    s to_i24 { I24::new_unchecked(ops::f64::round(s * 8_388_608.0) as i32) }
+    s to_i32 { ops::f64::round(s * 2_147_483_648.0) as i32 }
+    s to_i48 { I48::new_unchecked(ops::f64::round(s * 140_737_488_355_328.0) as i64) }
+    s to_i64 { ops::f64::round(s * 9_223_372_036_854_775_808.0) as i64 }
     s to_u8 { super::i8::to_u8(to_i8(s)) }
     s to_u16 { super::i16::to_u16(to_i16(s)) }
     s to_u24 { super::i24::to_u24(to_i24(s)) }
diff --git a/dasp_sample/src/ops.rs b/dasp_sample/src/ops.rs
index 81a5b5c7..6fbd5a20 100644
--- a/dasp_sample/src/ops.rs
+++ b/dasp_sample/src/ops.rs
@@ -3,6 +3,7 @@ pub mod f32 {
     /// Uses bit manipulation for initial guess, then 3 iterations for ~6-7 decimal places.
     /// Accuracy: ~6-7 decimal places
     #[cfg(not(feature = "std"))]
+    #[inline]
     pub fn sqrt(x: f32) -> f32 {
         if x < 0.0 {
             return f32::NAN;
@@ -31,6 +32,19 @@ pub mod f32 {
     pub fn sqrt(x: f32) -> f32 {
         x.sqrt()
     }
+
+    #[cfg(not(feature = "std"))]
+    #[inline]
+    pub fn round(x: f32) -> f32 {
+        // Branchless rounding: copysign gives +0.5 for positive x, -0.5 for negative x
+        // This shifts the value toward zero before truncation, achieving proper rounding
+        (x + 0.5_f32.copysign(x)) as i64 as f32
+    }
+    #[cfg(feature = "std")]
+    #[inline]
+    pub fn round(x: f32) -> f32 {
+        x.round()
+    }
 }
 
 pub mod f64 {
@@ -38,6 +52,7 @@ pub mod f64 {
     /// Uses bit manipulation for initial guess, then 4 iterations for ~14-15 decimal places.
     /// Accuracy: ~14-15 decimal places
     #[cfg(not(feature = "std"))]
+    #[inline]
     pub fn sqrt(x: f64) -> f64 {
         if x < 0.0 {
             return f64::NAN;
@@ -66,4 +81,17 @@ pub mod f64 {
     pub fn sqrt(x: f64) -> f64 {
         x.sqrt()
     }
+
+    #[cfg(not(feature = "std"))]
+    #[inline]
+    pub fn round(x: f64) -> f64 {
+        // Branchless rounding: copysign gives +0.5 for positive x, -0.5 for negative x
+        // This shifts the value toward zero before truncation, achieving proper rounding
+        (x + 0.5_f64.copysign(x)) as i64 as f64
+    }
+    #[cfg(feature = "std")]
+    #[inline]
+    pub fn round(x: f64) -> f64 {
+        x.round()
+    }
 }
diff --git a/dasp_sample/tests/conv.rs b/dasp_sample/tests/conv.rs
index b442d246..fee081d1 100644
--- a/dasp_sample/tests/conv.rs
+++ b/dasp_sample/tests/conv.rs
@@ -479,11 +479,11 @@ tests!(u64 {
 });
 
 tests!(f32 {
-    to_i8  { -1.0, -128; 0.0, 0; }
-    to_i16 { -1.0, -32_768; 0.0, 0; }
-    to_i24 { -1.0, -8_388_608; 0.0, 0; }
-    to_i32 { -1.0, -2_147_483_648; 0.0, 0; }
-    to_i48 { -1.0, -140_737_488_355_328; 0.0, 0; }
+    to_i8  { -1.0, -128; 0.0, 0; 0.1, 13; 0.004, 1; -0.004, -1; 0.003, 0; }
+    to_i16 { -1.0, -32_768; 0.0, 0; 0.1, 3277; 0.00002, 1; 0.00001, 0; }
+    to_i24 { -1.0, -8_388_608; 0.0, 0; 0.1, 838861; 0.0000001, 1; -0.0000001, -1; 0.00000005, 0; }
+    to_i32 { -1.0, -2_147_483_648; 0.0, 0; 0.0000000004, 1; -0.0000000004, -1; 0.0000000002, 0; }
+    to_i48 { -1.0, -140_737_488_355_328; 0.0, 0; 0.000000000000006, 1; -0.000000000000006, -1; 0.000000000000003, 0; }
     to_i64 { -1.0, -9_223_372_036_854_775_808; 0.0, 0; }
     to_u8  { -1.0, 0; 0.0, 128; }
     to_u16 { -1.0, 0; 0.0, 32_768; }
@@ -495,12 +495,12 @@ tests!(f32 {
 });
 
 tests!(f64 {
-    to_i8  { -1.0, -128; 0.0, 0; }
-    to_i16 { -1.0, -32_768; 0.0, 0; }
-    to_i24 { -1.0, -8_388_608; 0.0, 0; }
-    to_i32 { -1.0, -2_147_483_648; 0.0, 0; }
-    to_i48 { -1.0, -140_737_488_355_328; 0.0, 0; }
-    to_i64 { -1.0, -9_223_372_036_854_775_808; 0.0, 0; }
+    to_i8  { -1.0, -128; 0.0, 0; 0.1, 13; 0.007, 1; -0.004, -1; 0.003, 0; }
+    to_i16 { -1.0, -32_768; 0.0, 0; 0.1, 3277; 0.00002, 1; -0.00002, -1; 0.00001, 0; }
+    to_i24 { -1.0, -8_388_608; 0.0, 0; 0.1, 838861; 0.0000001, 1; -0.0000001, -1; 0.00000005, 0; }
+    to_i32 { -1.0, -2_147_483_648; 0.0, 0; 0.1, 214748365; 0.0000000004, 1; -0.0000000004, -1; 0.0000000002, 0; }
+    to_i48 { -1.0, -140_737_488_355_328; 0.0, 0; 0.1, 14073748835533; 0.000000000000006, 1; -0.000000000000006, -1; 0.000000000000003, 0; }
+    to_i64 { -1.0, -9_223_372_036_854_775_808; 0.0, 0; 0.1, 922337203685477632; }
     to_u8  { -1.0, 0; 0.0, 128; }
     to_u16 { -1.0, 0; 0.0, 32_768; }
     to_u24 { -1.0, 0; 0.0, 8_388_608; }