feat: add a resampling feature

pyproject.toml

@@ -20,6 +20,7 @@ websockets = "^10.3"
 tqdm = "*"
 torchaudio = "^0.12.1"
 PySoundFile = {version = "^0.9.0.post1", platform = "windows"}
+scipy = "^1.9.3"
 
 
 [tool.poetry.group.dev.dependencies]

whispering/resample.py

@@ -0,0 +1,24 @@
+#!/usr/bin/env python3
+import numpy as np
+from scipy.interpolate import interp1d
+from whisper.audio import SAMPLE_RATE
+
+
+def resample(x: np.ndarray, source_sr: int) -> np.ndarray:
+    """Resample a numpy array to match the SAMPLE_RATE
+    of Whisper (16 000 Hz).
+
+    Args:
+        x (np.ndarray): Source numpy array
+        source_sr (int): Source sample rate
+
+    Returns:
+        np.ndarray: The resampled array
+    """
+
+    if source_sr == SAMPLE_RATE:
+        return x
+    factor = source_sr / SAMPLE_RATE
+    n = int(np.ceil(x.size / factor))
+    f = interp1d(np.linspace(0, 1, x.size), x, "linear")
+    return f(np.linspace(0, 1, n))

whispering/schema.py

@@ -49,6 +49,7 @@ class Context(BaseModel, arbitrary_types_allowed=True):
     max_nospeech_skip: int
 
     data_type: str = "float32"
+    source_sample_rate: int = 16000
 
 
 class ParsedChunk(BaseModel):

whispering/serve.py

@@ -11,6 +11,7 @@
 
 from whispering.schema import CURRENT_PROTOCOL_VERSION, Context
 from whispering.transcriber import WhisperStreamingTranscriber
+from whispering.resample import resample
 
 logger = getLogger(__name__)
 
@@ -77,6 +78,7 @@ async def serve_with_websocket_main(websocket):
             )
             return
         audio = np.frombuffer(message, dtype=np.dtype(ctx.data_type)).astype(np.float32)
+        audio = resample(audio, ctx.source_sample_rate)
         for chunk in g_wsp.transcribe(
             audio=audio,  # type: ignore
             ctx=ctx,