Fix StreamVAD audio slicing

.gitignore

@@ -1,2 +1,3 @@
 output
-__pycache__
+__pycache__
+.vscode

SileroOrt.py

@@ -11,8 +11,6 @@ class SileroOrt:
         self.sr = 16000
         self.hidden_size = 128
         self.context_size = 64 if self.sr == 16000 else 32
-        self.context = np.zeros((self.batch_size, self.context_size), dtype=np.float32)
-        self.state = np.zeros((2, self.batch_size, self.hidden_size), dtype=np.float32)
         self.num_samples = 512 if self.sr == 16000 else 256
 
         self.model = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])

StreamVAD.py

@@ -41,11 +41,14 @@ class StreamVAD:
         # 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 : (i + 1) * self.model.num_samples]
+            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)
 
             # is speech

gradio_app.py

@@ -4,7 +4,9 @@ from dataclasses import dataclass, field
 
 
 vad = StreamVAD(
-    'silero_vad.onnx'
+    'silero_vad.onnx',
+    sensitivity=0.5,
+    silence_ms=200
 )
 
 @dataclass
@@ -12,7 +14,7 @@ class AppState:
     history: list = field(default_factory=list)
 
 
-def process_audio(audio, chatbot, state):
+def process_audio(audio, state):
     # print(audio)
     # audio is a tuple of (sample_rate, numpy int16 array)
     sr, audio_data = audio
@@ -21,13 +23,12 @@ def process_audio(audio, chatbot, state):
             state.history.append(
                 gr.ChatMessage(role='user', content=gr.Audio(
                     label=f"{result['start_ts']} - {result['end_ts']}",
-                    value=(vad.model.sr, result['audio']),
+                    value=(sr, result['audio']),
                     waveform_options=gr.WaveformOptions(show_recording_waveform=False),
                     editable=False
                     )
                 ),
             )
-    
 
     return state.history
 
@@ -42,6 +43,6 @@ with gr.Blocks() as demo:
         input_audio = gr.Audio(sources=['microphone'], type='numpy', streaming=True)
 
     # streaming process
-    input_audio.stream(fn=process_audio, inputs=[input_audio, chatbot, state], outputs=[chatbot])
+    input_audio.stream(fn=process_audio, inputs=[input_audio, state], outputs=[chatbot])
 
 demo.launch(debug=True)