diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index 0c85eb044..22585dc9e 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -28,6 +28,18 @@ jobs:
         run: swift test --parallel --num-workers $(sysctl -n hw.ncpu)
         timeout-minutes: 20
 
+  build-macos-x86_64:
+    name: Build Swift Package (macOS x86_64 cross-compile)
+    runs-on: macos-15
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v5
+
+      # Intel macs have no Swift Float16; this guards against unguarded
+      # fp16 usage breaking universal builds (issue #875).
+      - name: Build package (x86_64)
+        run: swift build --arch x86_64
+
   build-ios:
     name: Build (iOS)
     runs-on: macos-15
diff --git a/Sources/FluidAudio/ASR/Paraformer/ParaformerManager.swift b/Sources/FluidAudio/ASR/Paraformer/ParaformerManager.swift
index 3f0706eee..c0cfc14bf 100644
--- a/Sources/FluidAudio/ASR/Paraformer/ParaformerManager.swift
+++ b/Sources/FluidAudio/ASR/Paraformer/ParaformerManager.swift
@@ -474,11 +474,15 @@ public actor ParaformerManager {
                 out.append(Array(UnsafeBufferPointer(start: p + t * frameStride, count: dim)))
             }
         } else {
-            let p = arr.dataPointer.assumingMemoryBound(to: Float16.self)
+            // Address fp16 storage as raw bit patterns: Swift's `Float16` is
+            // unavailable on macOS x86_64.
+            let p = arr.dataPointer.assumingMemoryBound(to: UInt16.self)
             for t in 0..<count {
                 var r = [Float](repeating: 0, count: dim)
-                let base = t * frameStride
-                for d in 0..<dim { r[d] = Float(p[base + d]) }
+                r.withUnsafeMutableBufferPointer { dst in
+                    Float16Conversion.toFloat32(
+                        src: p + t * frameStride, dst: dst.baseAddress!, count: dim)
+                }
                 out.append(r)
             }
         }
diff --git a/Sources/FluidAudio/ASR/Shared/LogitsArgmax.swift b/Sources/FluidAudio/ASR/Shared/LogitsArgmax.swift
index 40a604468..72524b6a1 100644
--- a/Sources/FluidAudio/ASR/Shared/LogitsArgmax.swift
+++ b/Sources/FluidAudio/ASR/Shared/LogitsArgmax.swift
@@ -32,22 +32,12 @@ enum LogitsArgmax {
             run(logits.dataPointer.assumingMemoryBound(to: Float32.self))
         } else {
             let count = frames * frameStride
-            let src = logits.dataPointer.assumingMemoryBound(to: Float16.self)
+            // Address fp16 storage as raw bit patterns: Swift's `Float16` is
+            // unavailable on macOS x86_64.
+            let src = logits.dataPointer.assumingMemoryBound(to: UInt16.self)
             var buf = [Float](repeating: 0, count: count)
-            // The vImage buffers must not outlive the pointers they wrap, so the
-            // convert + argmax both run inside the withUnsafe scope.
             buf.withUnsafeMutableBufferPointer { dst in
-                var srcBuf = vImage_Buffer(
-                    data: UnsafeMutableRawPointer(mutating: src),
-                    height: 1,
-                    width: vImagePixelCount(count),
-                    rowBytes: count * MemoryLayout<Float16>.size)
-                var dstBuf = vImage_Buffer(
-                    data: dst.baseAddress!,
-                    height: 1,
-                    width: vImagePixelCount(count),
-                    rowBytes: count * MemoryLayout<Float>.size)
-                vImageConvert_Planar16FtoPlanarF(&srcBuf, &dstBuf, 0)
+                Float16Conversion.toFloat32(src: src, dst: dst.baseAddress!, count: count)
                 run(dst.baseAddress!)
             }
         }
diff --git a/Sources/FluidAudio/Shared/Float16Conversion.swift b/Sources/FluidAudio/Shared/Float16Conversion.swift
new file mode 100644
index 000000000..6ffdc6529
--- /dev/null
+++ b/Sources/FluidAudio/Shared/Float16Conversion.swift
@@ -0,0 +1,37 @@
+import Accelerate
+
+/// Portable fp16 ↔ fp32 buffer conversions.
+///
+/// Swift's `Float16` is unavailable on macOS x86_64, so half-precision
+/// buffers are addressed as `UInt16` bit patterns and converted with
+/// vImage's Planar16F routines, which compile on every architecture.
+enum Float16Conversion {
+
+    /// Widen `count` half-precision values (as raw bit patterns) to Float32.
+    static func toFloat32(
+        src: UnsafePointer<UInt16>, dst: UnsafeMutablePointer<Float>, count: Int
+    ) {
+        var srcBuf = vImage_Buffer(
+            data: UnsafeMutableRawPointer(mutating: src), height: 1,
+            width: vImagePixelCount(count),
+            rowBytes: count * MemoryLayout<UInt16>.stride)
+        var dstBuf = vImage_Buffer(
+            data: dst, height: 1, width: vImagePixelCount(count),
+            rowBytes: count * MemoryLayout<Float>.stride)
+        vImageConvert_Planar16FtoPlanarF(&srcBuf, &dstBuf, 0)
+    }
+
+    /// Narrow `count` Float32 values to half precision (as raw bit patterns).
+    static func fromFloat32(
+        src: UnsafePointer<Float>, dst: UnsafeMutablePointer<UInt16>, count: Int
+    ) {
+        var srcBuf = vImage_Buffer(
+            data: UnsafeMutableRawPointer(mutating: src), height: 1,
+            width: vImagePixelCount(count),
+            rowBytes: count * MemoryLayout<Float>.stride)
+        var dstBuf = vImage_Buffer(
+            data: dst, height: 1, width: vImagePixelCount(count),
+            rowBytes: count * MemoryLayout<UInt16>.stride)
+        vImageConvert_PlanarFtoPlanar16F(&srcBuf, &dstBuf, 0)
+    }
+}
diff --git a/Sources/FluidAudio/TTS/NeuTts/NeuTtsSynthesizer.swift b/Sources/FluidAudio/TTS/NeuTts/NeuTtsSynthesizer.swift
index b5a7ed2b8..39114e2c7 100644
--- a/Sources/FluidAudio/TTS/NeuTts/NeuTtsSynthesizer.swift
+++ b/Sources/FluidAudio/TTS/NeuTts/NeuTtsSynthesizer.swift
@@ -167,10 +167,10 @@ struct NeuTtsSynthesizer {
                 state.withMultiArray(for: name) { array in
                     array.withUnsafeMutableBytes { rawBuffer, _ in
                         guard let dstBase = rawBuffer.baseAddress else { return }
-                        let dst = dstBase.assumingMemoryBound(to: Float16.self)
-                        for i in 0..<layerElements {
-                            dst[i] = Float16(base[i])
-                        }
+                        // Address fp16 storage as raw bit patterns: Swift's
+                        // `Float16` is unavailable on macOS x86_64.
+                        let dst = dstBase.assumingMemoryBound(to: UInt16.self)
+                        Float16Conversion.fromFloat32(src: base, dst: dst, count: layerElements)
                     }
                 }
             }
@@ -227,8 +227,10 @@ struct NeuTtsSynthesizer {
                 dst.baseAddress!.update(from: ptr, count: count)
             }
         case .float16:
-            let ptr = array.dataPointer.assumingMemoryBound(to: Float16.self)
-            for i in 0..<count { out[i] = Float(ptr[i]) }
+            let ptr = array.dataPointer.assumingMemoryBound(to: UInt16.self)
+            out.withUnsafeMutableBufferPointer { dst in
+                Float16Conversion.toFloat32(src: ptr, dst: dst.baseAddress!, count: count)
+            }
         default:
             throw SynthesisError.missingOutput("unexpected dataType \(array.dataType.rawValue)")
         }
diff --git a/Tests/FluidAudioTests/Shared/Float16ConversionTests.swift b/Tests/FluidAudioTests/Shared/Float16ConversionTests.swift
new file mode 100644
index 000000000..27660d04f
--- /dev/null
+++ b/Tests/FluidAudioTests/Shared/Float16ConversionTests.swift
@@ -0,0 +1,61 @@
+import XCTest
+
+@testable import FluidAudio
+
+/// Uses hard-coded IEEE-754 binary16 bit patterns (not `Float16`) so the
+/// test itself compiles on macOS x86_64.
+final class Float16ConversionTests: XCTestCase {
+
+    private let cases: [(bits: UInt16, value: Float)] = [
+        (0x0000, 0.0),
+        (0x3800, 0.5),
+        (0x3C00, 1.0),
+        (0xC000, -2.0),
+        (0x4248, 3.140625),
+        (0x7BFF, 65504.0),  // largest finite half
+        (0x0001, Float(5.960464477539063e-08)),  // smallest subnormal
+    ]
+
+    func testToFloat32KnownBitPatterns() {
+        let src = cases.map { $0.bits }
+        var dst = [Float](repeating: .nan, count: src.count)
+        src.withUnsafeBufferPointer { s in
+            dst.withUnsafeMutableBufferPointer { d in
+                Float16Conversion.toFloat32(src: s.baseAddress!, dst: d.baseAddress!, count: s.count)
+            }
+        }
+        for (i, expected) in cases.map({ $0.value }).enumerated() {
+            XCTAssertEqual(dst[i], expected, "index \(i)")
+        }
+    }
+
+    func testFromFloat32KnownBitPatterns() {
+        let src = cases.map { $0.value }
+        var dst = [UInt16](repeating: 0xFFFF, count: src.count)
+        src.withUnsafeBufferPointer { s in
+            dst.withUnsafeMutableBufferPointer { d in
+                Float16Conversion.fromFloat32(src: s.baseAddress!, dst: d.baseAddress!, count: s.count)
+            }
+        }
+        for (i, expected) in cases.map({ $0.bits }).enumerated() {
+            XCTAssertEqual(dst[i], expected, "index \(i)")
+        }
+    }
+
+    func testRoundTripIsExactForHalfRepresentableValues() {
+        let values: [Float] = [0.0, -0.125, 0.25, 1.5, -3.0, 1024.0, -65504.0]
+        var bits = [UInt16](repeating: 0, count: values.count)
+        var back = [Float](repeating: .nan, count: values.count)
+        values.withUnsafeBufferPointer { s in
+            bits.withUnsafeMutableBufferPointer { d in
+                Float16Conversion.fromFloat32(src: s.baseAddress!, dst: d.baseAddress!, count: s.count)
+            }
+        }
+        bits.withUnsafeBufferPointer { s in
+            back.withUnsafeMutableBufferPointer { d in
+                Float16Conversion.toFloat32(src: s.baseAddress!, dst: d.baseAddress!, count: s.count)
+            }
+        }
+        XCTAssertEqual(back, values)
+    }
+}