diff --git a/pkgs/by-name/wh/whisper-cpp/package.nix b/pkgs/by-name/wh/whisper-cpp/package.nix
index e54b0c4f5bd8..699e1cab4199 100644
--- a/pkgs/by-name/wh/whisper-cpp/package.nix
+++ b/pkgs/by-name/wh/whisper-cpp/package.nix
@@ -8,6 +8,7 @@
   SDL2,
   wget,
   which,
+  ffmpeg,
   autoAddDriverRunpath,
   makeWrapper,
   nix-update-script,
@@ -29,10 +30,13 @@
   vulkan-loader,
 
   withSDL ? true,
+
+  withFFmpegSupport ? false,
 }:
 
 assert metalSupport -> stdenv.hostPlatform.isDarwin;
 assert coreMLSupport -> stdenv.hostPlatform.isDarwin;
+assert withFFmpegSupport -> stdenv.hostPlatform.isLinux;
 
 let
   # It's necessary to consistently use backendStdenv when building with CUDA support,
@@ -108,6 +112,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
 
   buildInputs =
     optional withSDL SDL2
+    ++ optional withFFmpegSupport ffmpeg
     ++ optionals cudaSupport cudaBuildInputs
     ++ optionals rocmSupport rocmBuildInputs
     ++ optionals vulkanSupport vulkanBuildInputs;
@@ -122,6 +127,9 @@ effectiveStdenv.mkDerivation (finalAttrs: {
     (cmakeBool "GGML_NATIVE" false)
     (cmakeBool "BUILD_SHARED_LIBS" (!effectiveStdenv.hostPlatform.isStatic))
   ]
+  ++ optionals effectiveStdenv.hostPlatform.isLinux [
+    (cmakeBool "WHISPER_FFMPEG" withFFmpegSupport)
+  ]
   ++ optionals (effectiveStdenv.hostPlatform.isx86 && !effectiveStdenv.hostPlatform.isStatic) [
     (cmakeBool "GGML_BACKEND_DL" true)
     (cmakeBool "GGML_CPU_ALL_VARIANTS" true)
@@ -156,6 +164,10 @@ effectiveStdenv.mkDerivation (finalAttrs: {
 
     wrapProgram "$out/bin/whisper-cpp-download-ggml-model" \
       --prefix PATH : ${lib.makeBinPath [ wget ]}
+  ''
+  + lib.optionalString withFFmpegSupport ''
+    wrapProgram "$out/bin/whisper-server" \
+      --prefix PATH : ${lib.makeBinPath [ ffmpeg ]}
   '';
 
   requiredSystemFeatures = optionals rocmSupport [ "big-parallel" ]; # rocmSupport multiplies build time by the number of GPU targets, which takes arround 30 minutes on a 16-cores system to build