Replace torchaudio with soundfile in python-api-examples (#765)

python-api-examples/speaker-identification-with-vad-non-streaming-asr.py

@@ -65,7 +65,7 @@ from typing import Dict, List, Tuple
 
 import numpy as np
 import sherpa_onnx
-import torchaudio
+import soundfile as sf
 
 try:
     import sounddevice as sd
@@ -357,8 +357,14 @@ def load_speaker_file(args) -> Dict[str, List[str]]:
 
 
 def load_audio(filename: str) -> Tuple[np.ndarray, int]:
-    samples, sample_rate = torchaudio.load(filename)
-    return samples[0].contiguous().numpy(), sample_rate
+    data, sample_rate = sf.read(
+        filename,
+        always_2d=True,
+        dtype="float32",
+    )
+    data = data[:, 0]  # use only the first channel
+    samples = np.ascontiguousarray(data)
+    return samples, sample_rate
 
 
 def compute_speaker_embedding(

python-api-examples/speaker-identification-with-vad.py

@@ -60,7 +60,7 @@ from typing import Dict, List, Tuple
 
 import numpy as np
 import sherpa_onnx
-import torchaudio
+import soundfile as sf
 
 try:
     import sounddevice as sd
@@ -160,8 +160,14 @@ def load_speaker_file(args) -> Dict[str, List[str]]:
 
 
 def load_audio(filename: str) -> Tuple[np.ndarray, int]:
-    samples, sample_rate = torchaudio.load(filename)
-    return samples[0].contiguous().numpy(), sample_rate
+    data, sample_rate = sf.read(
+        filename,
+        always_2d=True,
+        dtype="float32",
+    )
+    data = data[:, 0]  # use only the first channel
+    samples = np.ascontiguousarray(data)
+    return samples, sample_rate
 
 
 def compute_speaker_embedding(

python-api-examples/speaker-identification.py

@@ -52,7 +52,7 @@ from typing import Dict, List, Tuple
 
 import numpy as np
 import sherpa_onnx
-import torchaudio
+import soundfile as sf
 
 try:
     import sounddevice as sd
@@ -145,8 +145,14 @@ def load_speaker_file(args) -> Dict[str, List[str]]:
 
 
 def load_audio(filename: str) -> Tuple[np.ndarray, int]:
-    samples, sample_rate = torchaudio.load(filename)
-    return samples[0].contiguous().numpy(), sample_rate
+    data, sample_rate = sf.read(
+        filename,
+        always_2d=True,
+        dtype="float32",
+    )
+    data = data[:, 0]  # use only the first channel
+    samples = np.ascontiguousarray(data)
+    return samples, sample_rate
 
 
 def compute_speaker_embedding(