diff --git a/speech.py b/speech.py
index 0ea970e..09c320f 100755
--- a/speech.py
+++ b/speech.py
@@ -602,6 +602,8 @@ def build_ffmpeg_args(response_format, input_format, sample_rate):
         ffmpeg_args.extend(["-f", "flac", "-c:a", "flac"])
     elif response_format == "wav":
         ffmpeg_args.extend(["-f", "wav", "-c:a", "pcm_s16le"])
+    elif response_format == "webm":
+        ffmpeg_args.extend(["-f", "webm", "-c:a", "libopus"])
     elif response_format == "pcm": # even though pcm is technically 'raw', we still use ffmpeg to adjust the speed
         ffmpeg_args.extend(["-f", "s16le", "-c:a", "pcm_s16le"])
 
@@ -749,6 +751,8 @@ async def generate_speech(request: GenerateSpeechRequest):
         media_type = "audio/x-flac"
     elif response_format == "wav":
         media_type = "audio/wav"
+    elif response_format == "webm":
+        media_type = "audio/webm;codecs=opus"
     elif response_format == "pcm":
         if model == 'tts-1': # piper
             media_type = "audio/pcm;rate=22050"