align with silero-vad-axera 0.1.2 (axmodel-only)

README.md

@@ -6,22 +6,25 @@ license: mit
 流式语音端点识别
 
 
-## 安装依赖
-
-```
-sudo apt install libsndfile1
-
-pip install -r requirements.txt
-```
+## 安装依赖
+
+本仓库只支持 **axmodel（NPU）推理**（AX650 / AX630C），不再支持 onnx CPU 后端。
+
+```
+sudo apt install libsndfile1
+
+pip install -r requirements.txt
+```
 
 ## Demo
 
 ### CLI
 
-```
-python main.py --input demo.wav --output_dir output --model silero_vad.onnx
-```
-被分段的语音后保存在output目录中
+```
+python main.py --input demo.wav --backend ax650 --output_dir output   # AX650 板
+python main.py --input demo.wav --backend ax630c --output_dir output  # AX620E/AX630C 板
+```
+被分段的语音后保存在output目录中
 
 
 ### Gradio
@@ -63,4 +66,4 @@ result的格式为:
 }
 ```
 
-时间戳的格式可通过StreamVAD.datetime_format设置
+时间戳的格式可通过StreamVAD.datetime_format设置

StreamVAD.py

@@ -3,7 +3,7 @@ import numpy as np
 from datetime import datetime, timedelta
 
 
-class StreamVAD:
+class StreamVAD:
     def __init__(self, backend='ax650',
                         sensitivity=0.5,
                         silence_ms=200,
@@ -17,10 +17,13 @@ class StreamVAD:
         datetime_format: format of datetime in return data
         '''
 
-        self.model = load_silero_vad(backend)
-        self.sensitivity = sensitivity
-        self.silence_ms = silence_ms
-        self.datetime_format = datetime_format
+        self.model = load_silero_vad(backend)
+        # axmodel 为 16k 静态图：采样率 16000、每帧 512 样本
+        self.sr = self.model.sample_rates[0]
+        self.num_samples = 512
+        self.sensitivity = sensitivity
+        self.silence_ms = silence_ms
+        self.datetime_format = datetime_format
 
         self.reset()
 
@@ -37,19 +40,22 @@ class StreamVAD:
         self.model.reset_states()
 
     
-    def run(self, audio: np.ndarray, sr: int = 16000):
-        # record datetime
-        cur_ts = datetime.now()
-
-        # freq scale
-        freq_scale = int(sr / self.model.sr)
-
-        # inference
-        speech_probs = self.model.audio_forward(audio, sr)[0]
-
-        for i, prob in enumerate(speech_probs):
-            audio_slice = audio[i * self.model.num_samples * freq_scale : (i + 1) * self.model.num_samples * freq_scale]
-            ts = cur_ts.strftime(self.datetime_format)
+    def run(self, audio: np.ndarray, sr: int = 16000):
+        audio = np.asarray(audio, dtype=np.float32)
+        # record datetime
+        cur_ts = datetime.now()
+
+        # freq scale
+        freq_scale = int(sr / self.sr)
+
+        # inference
+        speech_probs = self.model.audio_forward(audio, sr)[0]
+
+        for i, prob in enumerate(speech_probs):
+            audio_slice = audio[i * self.num_samples * freq_scale : (i + 1) * self.num_samples * freq_scale]
+            if len(audio_slice) < self.num_samples * freq_scale:
+                audio_slice = np.pad(audio_slice, (0, self.num_samples * freq_scale - len(audio_slice)))
+            ts = cur_ts.strftime(self.datetime_format)
 
             # is speech
             if prob > 1 - self.sensitivity:
@@ -66,7 +72,7 @@ class StreamVAD:
                     self.silence_count += 1
 
                     # exceed silence limit
-                    if 1000 * self.silence_count * self.model.num_samples / self.model.sr > self.silence_ms:
+                    if 1000 * self.silence_count * self.num_samples / self.sr > self.silence_ms:
                         # return audio segment
                         self.return_data['end_ts'] = ts
                         self.return_data['audio'] = np.concatenate(self.vad_data_list, axis=-1)
@@ -78,4 +84,4 @@ class StreamVAD:
                         self.vad_data_list.append(audio_slice)
 
             # timestamp
-            cur_ts += timedelta(seconds=self.model.num_samples / self.model.sr)
+            cur_ts += timedelta(seconds=self.num_samples / self.sr)

main.py

@@ -8,7 +8,7 @@ import soundfile as sf
 def get_args():
     parser = argparse.ArgumentParser()
     parser.add_argument('--input', type=str, required=True, help='Input audio file')
-    parser.add_argument('--backend', type=str, default='ax650', choices=['ax650', 'ax630c', 'onnx'])
+    parser.add_argument('--backend', type=str, default='ax650', choices=['ax650', 'ax630c'])
     parser.add_argument('--output_dir', type=str, default='output', help='Output audio dir')
     return parser.parse_args()
 
@@ -21,15 +21,15 @@ def main():
                     sensitivity=0.5,
                     silence_ms=200)
     
-    audio, _ = librosa.load(args.input, sr=vad.model.sr, mono=True)
+    audio, _ = librosa.load(args.input, sr=vad.sr, mono=True)
     i = 0
-    for result in vad.run(audio, vad.model.sr):
+    for result in vad.run(audio, vad.sr):
         if result:
             print(result)
             filename = os.path.join(args.output_dir, f"{i}.wav")
-            sf.write(filename, result['audio'], samplerate=vad.model.sr)
+            sf.write(filename, result['audio'], samplerate=vad.sr)
             i += 1
 
 
 if __name__ == '__main__':
-    main()
+    main()

requirements.txt

@@ -1,6 +1,5 @@
-onnxruntime>=1.16.1
-librosa==0.9.1
-numpy<2.0
-soundfile
-axengine @ https://github.com/AXERA-TECH/pyaxengine/releases/download/0.1.3.rc2/axengine-0.1.3-py3-none-any.whl
-silero_vad_axera
+librosa==0.9.1
+numpy<2.0
+soundfile
+axengine @ https://github.com/AXERA-TECH/pyaxengine/releases/download/0.1.3.rc2/axengine-0.1.3-py3-none-any.whl
+silero-vad-axera>=0.1.2