diff --git a/pyroomacoustics/libroom_src/libroom.cpp b/pyroomacoustics/libroom_src/libroom.cpp
index 300a5395..444c6be2 100644
--- a/pyroomacoustics/libroom_src/libroom.cpp
+++ b/pyroomacoustics/libroom_src/libroom.cpp
@@ -282,10 +282,7 @@ PYBIND11_MODULE(libroom, m) {
   m.def("dist_line_point", &dist_line_point,
         "Computes the distance between a point and an infinite line");
 
-  m.def("rir_builder", &rir_builder, "RIR builder",
-        py::call_guard<py::gil_scoped_release>());
-  m.def("delay_sum", &delay_sum, "Delay and sum",
-        py::call_guard<py::gil_scoped_release>());
-  m.def("fractional_delay", &fractional_delay, "Fractional delays",
-        py::call_guard<py::gil_scoped_release>());
+  m.def("rir_builder", &rir_builder, "RIR builder");
+  m.def("delay_sum", &delay_sum, "Delay and sum");
+  m.def("fractional_delay", &fractional_delay, "Fractional delays");
 }
diff --git a/pyroomacoustics/libroom_src/rir_builder.cpp b/pyroomacoustics/libroom_src/rir_builder.cpp
index 5b599aa3..36cf0821 100644
--- a/pyroomacoustics/libroom_src/rir_builder.cpp
+++ b/pyroomacoustics/libroom_src/rir_builder.cpp
@@ -44,11 +44,12 @@ constexpr T get_pi() {
 template <class T>
 void threaded_rir_builder_impl(
     py::array_t<T, py::array::c_style | py::array::forcecast> rir,
-    const py::array_t<T, py::array::c_style | py::array::forcecast> time,
-    const py::array_t<T, py::array::c_style | py::array::forcecast> alpha,
+    const py::array_t<T, py::array::c_style | py::array::forcecast> &time,
+    const py::array_t<T, py::array::c_style | py::array::forcecast> &alpha,
     const py::array_t<int, py::array::c_style | py::array::forcecast>
-        visibility,
+        &visibility,
     int fs, size_t fdl, size_t lut_gran, size_t num_threads) {
+
   auto pi = get_pi<T>();
 
   // accessors for the arrays
@@ -110,6 +111,9 @@ void threaded_rir_builder_impl(
   std::vector<T> rir_out(num_threads * rir_len);
   size_t block_size = size_t(std::ceil(double(n_times) / double(num_threads)));
 
+  // relase the GIL from here on
+  py::gil_scoped_release release;
+
   // build the RIR
   ThreadPool pool(num_threads);
   std::vector<std::future<void>> results;
@@ -166,9 +170,10 @@ void threaded_rir_builder_impl(
   for (auto &&result : sum_results) result.get();
 }
 
-void rir_builder(py::buffer rir, const py::buffer time, const py::buffer alpha,
-                 const py::buffer visibility, int fs, size_t fdl,
+void rir_builder(py::buffer rir, const py::buffer &time, const py::buffer &alpha,
+                 const py::buffer &visibility, int fs, size_t fdl,
                  size_t lut_gran, size_t num_threads) {
+
   // dispatch to correct implementation depending on input type
   auto buf = pybind11::array::ensure(rir);
   if (py::isinstance<py::array_t<float>>(buf)) {
@@ -214,6 +219,9 @@ void threaded_delay_sum_impl(
   std::vector<T> out_buffers(num_threads * out_len, 0);
   size_t block_size = size_t(std::ceil(double(n_irs) / double(num_threads)));
 
+  // release the GIL from here on
+  py::gil_scoped_release release;
+
   // build the RIR
   ThreadPool pool(num_threads);
   std::vector<std::future<void>> results;
@@ -315,6 +323,9 @@ void threaded_fractional_delay_impl(
   // divide into equal size blocks for thread processing
   size_t block_size = size_t(std::ceil(double(n_times) / double(num_threads)));
 
+  // relase the GIL from here on
+  py::gil_scoped_release release;
+
   // build the RIR
   ThreadPool pool(num_threads);
   std::vector<std::future<void>> results;
diff --git a/pyroomacoustics/libroom_src/rir_builder.hpp b/pyroomacoustics/libroom_src/rir_builder.hpp
index f665123a..f4802a4c 100644
--- a/pyroomacoustics/libroom_src/rir_builder.hpp
+++ b/pyroomacoustics/libroom_src/rir_builder.hpp
@@ -7,8 +7,8 @@
 
 namespace py = pybind11;
 
-void rir_builder(py::buffer rir, const py::buffer time, const py::buffer alpha,
-                 const py::buffer visibility, int fs, size_t fdl,
+void rir_builder(py::buffer rir, const py::buffer &time, const py::buffer &alpha,
+                 const py::buffer &visibility, int fs, size_t fdl,
                  size_t lut_gran, size_t num_threads);
 
 void delay_sum(const py::buffer irs, const py::buffer delays, py::buffer output,
diff --git a/pyroomacoustics/tests/test_build_rir.py b/pyroomacoustics/tests/test_build_rir.py
index b1afc709..c28c7e14 100644
--- a/pyroomacoustics/tests/test_build_rir.py
+++ b/pyroomacoustics/tests/test_build_rir.py
@@ -285,7 +285,7 @@ def measure_runtime(dtype=np.float32, num_threads=4):
         td = np.round(tt).astype(np.int32)
         tf = (tt - td).astype(dtype)
         irs = np.zeros((tt.shape[0], fdl), dtype=dtype)
-        fractional_delay(irs, tf, 20, num_threads)
+        libroom.fractional_delay(irs, tf, 20, num_threads)
         irs *= alpha[:, None]
         libroom.delay_sum(irs, td, rir, num_threads)
     tock_2steps = (time.perf_counter() - tick) / n_repeat