Fix for silero vad v5. (#1065)

The network input is 64 + 512 samples instead of 512 samples for 16kHz.

Author: csukuangfj

sherpa-onnx/csrc/silero-vad-model.cc

@@ -74,9 +74,8 @@ class SileroVadModel::Impl {
   }
 
   bool IsSpeech(const float *samples, int32_t n) {
-    if (n != config_.silero_vad.window_size) {
-      SHERPA_ONNX_LOGE("n: %d != window_size: %d", n,
-                       config_.silero_vad.window_size);
+    if (n != WindowSize()) {
+      SHERPA_ONNX_LOGE("n: %d != window_size: %d", n, WindowSize());
       exit(-1);
     }
 
@@ -146,9 +145,11 @@ class SileroVadModel::Impl {
     return false;
   }
 
-  int32_t WindowSize() const { return config_.silero_vad.window_size; }
+  int32_t WindowShift() const { return config_.silero_vad.window_size; }
 
-  int32_t WindowShift() const { return WindowSize() - window_shift_; }
+  int32_t WindowSize() const {
+    return config_.silero_vad.window_size + window_overlap_;
+  }
 
   int32_t MinSilenceDurationSamples() const { return min_silence_samples_; }
 
@@ -177,9 +178,9 @@ class SileroVadModel::Impl {
 
       // 64 for 16kHz
       // 32 for 8kHz
-      window_shift_ = 64;
+      window_overlap_ = 64;
 
-      if (WindowSize() != 512) {
+      if (config_.silero_vad.window_size != 512) {
         SHERPA_ONNX_LOGE(
             "For silero_vad  v5, we require window_size to be 512 for 16kHz");
         exit(-1);
@@ -423,7 +424,7 @@ class SileroVadModel::Impl {
   int32_t temp_start_ = 0;
   int32_t temp_end_ = 0;
 
-  int32_t window_shift_ = 0;
+  int32_t window_overlap_ = 0;
 
   bool is_v5_ = false;
 };

sherpa-onnx/csrc/silero-vad-model.h

@@ -37,11 +37,12 @@ class SileroVadModel : public VadModel {
    */
   bool IsSpeech(const float *samples, int32_t n) override;
 
+  // For silero vad V4, it is WindowShift().
+  // For silero vad V5, it is WindowShift()+64 for 16kHz and
+  //                          WindowShift()+32 for 8kHz
   int32_t WindowSize() const override;
 
-  // For silero vad V4, it is WindowSize().
-  // For silero vad V5, it is WindowSize()-64 for 16kHz and
-  //                          WindowSize()-32 for 8kHz
+  // 512
   int32_t WindowShift() const override;
 
   int32_t MinSilenceDurationSamples() const override;

sherpa-onnx/csrc/voice-activity-detector.cc

@@ -44,13 +44,17 @@ class VoiceActivityDetector::Impl {
     // an extra buffer here
     last_.insert(last_.end(), samples, samples + n);
 
+    if (last_.size() < window_size) {
+      return;
+    }
+
     // Note: For v4, window_shift == window_size
     int32_t k =
         (static_cast<int32_t>(last_.size()) - window_size) / window_shift + 1;
     const float *p = last_.data();
     bool is_speech = false;
 
-    for (int32_t i = 0; i != k; ++i, p += window_shift) {
+    for (int32_t i = 0; i < k; ++i, p += window_shift) {
       buffer_.Push(p, window_shift);
       // NOTE(fangjun): Please don't use a very large n.
       bool this_window_is_speech = model_->IsSpeech(p, window_size);