fix: support uint8 websocket audio format

README.md

@@ -257,7 +257,7 @@ Accept raw PCM int16 audio from clients (useful for embedded devices):
 ```bash
 python3 run_server.py --port 9090 --backend faster_whisper --raw_pcm_input
 ```
-Audio is automatically normalized to float32 range [-1.0, 1.0].
+Audio is automatically normalized to float32 range [-1.0, 1.0]. Clients can also set `audio_format` in the initial websocket options to `float32` (default), `int16`, or `uint8`.
 
 ## Browser Extensions
 - Run the server with your desired backend as shown [here](https://github.com/collabora/WhisperLive?tab=readme-ov-file#running-the-server).

tests/test_server_extended.py

@@ -273,6 +273,16 @@ class TestTranscriptionServerGetAudio(unittest.TestCase):
         self.assertTrue(np.all(result >= -1.0))
         self.assertTrue(np.all(result <= 1.0))
 
+    def test_uint8_audio_format_normalizes_unsigned_pcm(self):
+        import numpy as np
+        ws = MagicMock()
+        self.server.audio_formats[ws] = "uint8"
+        pcm = np.array([0, 128, 255], dtype=np.uint8)
+        ws.recv.return_value = pcm.tobytes()
+        result = self.server.get_audio_from_websocket(ws)
+        expected = (pcm.astype(np.float32) - 128.0) / 128.0
+        np.testing.assert_array_almost_equal(result, expected)
+
     def test_raw_pcm_input_off_reads_float32(self):
         import numpy as np
         self.server.raw_pcm_input = False

whisper_live/server.py

@@ -178,6 +178,7 @@ class TranscriptionServer:
         self.single_model = False
         self.batch_config = None
         self.raw_pcm_input = False
+        self.audio_formats = {}
         self.segment_post_processor = None
 
     def initialize_client(
@@ -361,7 +362,11 @@ class TranscriptionServer:
         frame_data = websocket.recv()
         if frame_data == b"END_OF_AUDIO":
             return False
-        if self.raw_pcm_input:
+        audio_format = self.audio_formats.get(websocket)
+        if audio_format == "uint8":
+            audio_np = np.frombuffer(frame_data, dtype=np.uint8)
+            return (audio_np.astype(np.float32) - 128.0) / 128.0
+        if self.raw_pcm_input or audio_format == "int16":
             audio_np = np.frombuffer(frame_data, dtype=np.int16)
             return audio_np.astype(np.float32) / 32768.0
         return np.frombuffer(frame_data, dtype=np.float32)
@@ -378,6 +383,10 @@ class TranscriptionServer:
                 wl_metrics.track_connection_rejected(reason="full")
                 websocket.close()
                 return False  # Indicates that the connection should not continue
+            audio_format = options.get("audio_format", "float32")
+            if audio_format not in {"float32", "int16", "uint8"}:
+                raise ValueError(f"Unsupported audio_format: {audio_format}")
+            self.audio_formats[websocket] = audio_format
 
             if self.backend.is_tensorrt():
                 self.vad_detector = VoiceActivityDetector(frame_rate=self.RATE)
@@ -845,3 +854,4 @@ class TranscriptionServer:
             if hasattr(client, 'translation_thread') and client.translation_thread:
                 client.translation_thread.join(timeout=2.0)
             self.client_manager.remove_client(websocket)
+        self.audio_formats.pop(websocket, None)