diff --git a/src/ggml_graph.cpp b/src/ggml_graph.cpp
index 7bef2a6..cbcfcb8 100644
--- a/src/ggml_graph.cpp
+++ b/src/ggml_graph.cpp
@@ -81,6 +81,11 @@ Backend& global_backend() {
     return global_backend_locked();
 }
 
+int backend_thread_count() {
+    std::lock_guard<std::recursive_mutex> lock(g_backend_mutex);
+    return g_backend_threads;
+}
+
 void shutdown_backend() {
     // Free the process-global backend explicitly. Required for GPU backends: the
     // backend (and its gallocr's device buffer) must be released while the CUDA/
@@ -107,16 +112,26 @@ bool run_graph(size_t /*mem_bytes*/, int n_threads,
     std::lock_guard<std::recursive_mutex> lock(g_backend_mutex);
     Backend& be = global_backend_locked();
     // When no global override is set, honor the caller's per-call n_threads (the
-    // historical behavior, used by the unit tests). A positive global override
-    // already pinned the backend's thread count in global_backend().
+    // historical behavior, used by the unit tests) for THIS call only. The
+    // backend is process-global, so a count left behind would slow every later
+    // graph: a Silero VAD pass (1 thread per tiny chunk) used to leave the ASR
+    // decode single-threaded. The previous count is restored after the compute.
+    // A positive global override already pinned the backend's thread count in
+    // global_backend_locked() and wins over the per-call value.
     const int g = g_num_threads.load(std::memory_order_relaxed);
-    if (g <= 0 && n_threads > 0 && n_threads != g_backend_threads) {
+    const int prev_threads = g_backend_threads;
+    const bool scoped = g <= 0 && n_threads > 0 && n_threads != prev_threads;
+    if (scoped) {
         be.set_n_threads(n_threads);
         g_backend_threads = n_threads;
     }
 
     // Backend::compute builds the graph in a no_alloc context, allocates via the
     // persistent gallocr, pushes inputs AFTER alloc, computes, and reads back.
+    struct Restore {
+        Backend& be; bool on; int prev;
+        ~Restore() { if (on) { be.set_n_threads(prev); g_backend_threads = prev; } }
+    } restore{be, scoped, prev_threads};
     return be.compute(build, out);
 }
 
diff --git a/src/ggml_graph.hpp b/src/ggml_graph.hpp
index c3e9fef..a0cb33d 100644
--- a/src/ggml_graph.hpp
+++ b/src/ggml_graph.hpp
@@ -36,6 +36,11 @@ int  num_threads();  // current override (0 == unset)
 // if set, else the built-in default (8).
 int  effective_threads();
 
+// The thread count the process-global backend is using right now (0 before it
+// is created). A per-call `n_threads` in run_graph applies to that call only, so
+// this does not change across a call. For tests.
+int  backend_thread_count();
+
 // Gallocr buffer size (bytes) reserved for the most recent single-backend (CPU)
 // run_graph compute. Used by tests to assert that banded attention memory scales
 // O(T*window), not O(T^2). Reflects the high-water mark of the persistent
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index d015ced..1f13adc 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -7,6 +7,7 @@ function(pk_add_test name)
 endfunction()
 
 pk_add_test(test_smoke)
+pk_add_test(test_run_graph_threads)
 pk_add_test(test_nbest_json)
 pk_add_test(test_tdt_beam_core)
 pk_add_test(test_backend_device)
diff --git a/tests/test_run_graph_threads.cpp b/tests/test_run_graph_threads.cpp
new file mode 100644
index 0000000..1e06914
--- /dev/null
+++ b/tests/test_run_graph_threads.cpp
@@ -0,0 +1,61 @@
+// A per-call n_threads in run_graph must not change the thread count of the
+// process-global backend afterwards. Silero VAD asks for one thread per chunk;
+// before this was scoped, a VAD pass left the ASR decode that followed on one
+// thread. The Silero part runs only when PARAKEET_TEST_SILERO_GGUF is set.
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include <vector>
+
+#include "backend.hpp"
+#include "ggml.h"
+#include "ggml_graph.hpp"
+#include "silero_vad.hpp"
+
+static int failures = 0;
+#define CHECK(c) do { if (!(c)) { std::fprintf(stderr, "FAIL %s:%d %s\n", __FILE__, __LINE__, #c); ++failures; } } while (0)
+
+static bool tiny_graph(int n_threads) {
+    std::vector<float> out;
+    const float x[4] = {1, 2, 3, 4};
+    const bool ok = pk::run_graph(0, n_threads, [&](ggml_context* ctx) -> ggml_tensor* {
+        const int64_t ne[1] = {4};
+        ggml_tensor* in = pk::graph_input_tensor(ctx, GGML_TYPE_F32, 1, ne, x, sizeof(x));
+        return ggml_scale(ctx, in, 2.0f);
+    }, out);
+    return ok && out.size() == 4 && out[3] == 8.0f;
+}
+
+int main() {
+    pk::set_num_threads(0);  // no --threads
+    CHECK(tiny_graph(0));
+    const int before = pk::backend_thread_count();
+    CHECK(before == pk::effective_threads());
+
+    CHECK(tiny_graph(1));
+    CHECK(pk::backend_thread_count() == before);
+    CHECK(tiny_graph(before == 3 ? 2 : 3));
+    CHECK(pk::backend_thread_count() == before);
+
+    if (const char* gguf = std::getenv("PARAKEET_TEST_SILERO_GGUF")) {
+        std::string err;
+        auto vad = pk::SileroVad::load(gguf, &err);
+        CHECK(vad != nullptr);
+        if (vad) {
+            std::vector<float> pcm(16000, 0.0f);
+            CHECK(!vad->probabilities(pcm.data(), pcm.size(), 16000).empty());
+            CHECK(pk::backend_thread_count() == before);
+        }
+    }
+
+    // A global override still wins over the per-call count.
+    pk::set_num_threads(2);
+    CHECK(tiny_graph(1));
+    CHECK(pk::backend_thread_count() == 2);
+    pk::set_num_threads(0);
+
+    pk::shutdown_backend();
+    if (failures) return 1;
+    std::printf("ok\n");
+    return 0;
+}