Merge pull request #437 from boxerab/rawpcm

audio: add support for raw pcm input via server flag

run_server.py

@@ -84,6 +84,12 @@ if __name__ == "__main__":
         default=50,
         help='Maximum time in ms to wait for batch to fill (default: 50).'
     )
+    parser.add_argument(
+        '--raw_pcm_input',
+        action='store_true',
+        help='Expect raw PCM int16 audio from clients instead of float32. '
+             'Audio will be normalized to float32 range [-1.0, 1.0].'
+    )
     args = parser.parse_args()
 
     if args.backend == "tensorrt":
@@ -113,4 +119,5 @@ if __name__ == "__main__":
         batch_enabled=args.batch_inference,
         batch_max_size=args.batch_max_size,
         batch_window_ms=args.batch_window_ms,
+        raw_pcm_input=args.raw_pcm_input,
     )

tests/test_server_extended.py

@@ -169,6 +169,28 @@ class TestTranscriptionServerGetAudio(unittest.TestCase):
         result = self.server.get_audio_from_websocket(ws)
         np.testing.assert_array_almost_equal(result, audio)
 
+    def test_raw_pcm_input_normalizes_int16(self):
+        import numpy as np
+        self.server.raw_pcm_input = True
+        ws = MagicMock()
+        pcm = np.array([0, 16384, -16384, 32767], dtype=np.int16)
+        ws.recv.return_value = pcm.tobytes()
+        result = self.server.get_audio_from_websocket(ws)
+        expected = pcm.astype(np.float32) / 32768.0
+        np.testing.assert_array_almost_equal(result, expected)
+        self.assertTrue(result.dtype == np.float32)
+        self.assertTrue(np.all(result >= -1.0))
+        self.assertTrue(np.all(result <= 1.0))
+
+    def test_raw_pcm_input_off_reads_float32(self):
+        import numpy as np
+        self.server.raw_pcm_input = False
+        ws = MagicMock()
+        audio = np.array([0.5, -0.5], dtype=np.float32)
+        ws.recv.return_value = audio.tobytes()
+        result = self.server.get_audio_from_websocket(ws)
+        np.testing.assert_array_almost_equal(result, audio)
+
 
 class TestTranscriptionServerHandleNewConnection(unittest.TestCase):
     def setUp(self):

whisper_live/server.py

@@ -160,6 +160,7 @@ class TranscriptionServer:
         self.use_vad = True
         self.single_model = False
         self.batch_config = None
+        self.raw_pcm_input = False
 
     def initialize_client(
         self, websocket, options, faster_whisper_custom_model_path,
@@ -316,6 +317,9 @@ class TranscriptionServer:
         frame_data = websocket.recv()
         if frame_data == b"END_OF_AUDIO":
             return False
+        if self.raw_pcm_input:
+            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)
 
     def handle_new_connection(self, websocket, faster_whisper_custom_model_path,
@@ -431,7 +435,8 @@ class TranscriptionServer:
             cors_origins: Optional[str] = None,
             batch_enabled=False,
             batch_max_size=8,
-            batch_window_ms=50):
+            batch_window_ms=50,
+            raw_pcm_input=False):
         """
         Run the transcription server.
 
@@ -449,6 +454,7 @@ class TranscriptionServer:
                 to 50.
         """
         self.cache_path = cache_path
+        self.raw_pcm_input = raw_pcm_input
         self.client_manager = ClientManager(max_clients, max_connection_time)
         if faster_whisper_custom_model_path is not None and not os.path.exists(faster_whisper_custom_model_path):
             if "/" not in faster_whisper_custom_model_path: