From 6f10d1f07cf14cac477287c2ad513337554f7351 Mon Sep 17 00:00:00 2001
From: yujinqiu <yujinqiuustb@gmail.com>
Date: Mon, 16 Oct 2023 11:15:29 +0800
Subject: [PATCH] Add vad clear api for better performance

---
 .../com/k2fsa/sherpa/onnx/MainActivity.kt     |  2 +-
 sherpa-onnx/c-api/c-api.cc                    |  5 +++
 sherpa-onnx/c-api/c-api.h                     |  4 +++
 sherpa-onnx/csrc/voice-activity-detector.cc   |  2 ++
 sherpa-onnx/csrc/voice-activity-detector.h    |  1 +
 sherpa-onnx/jni/jni.cc                        | 10 ++++++
 swift-api-examples/SherpaOnnx.swift           |  4 +++
 swift-api-examples/generate-subtitles.swift   | 35 +++++++++----------
 8 files changed, 44 insertions(+), 19 deletions(-)

diff --git a/android/SherpaOnnxVad/app/src/main/java/com/k2fsa/sherpa/onnx/MainActivity.kt b/android/SherpaOnnxVad/app/src/main/java/com/k2fsa/sherpa/onnx/MainActivity.kt
index eb847b020..4d5ce7e74 100644
--- a/android/SherpaOnnxVad/app/src/main/java/com/k2fsa/sherpa/onnx/MainActivity.kt
+++ b/android/SherpaOnnxVad/app/src/main/java/com/k2fsa/sherpa/onnx/MainActivity.kt
@@ -161,9 +161,9 @@ class MainActivity : AppCompatActivity() {
                 val samples = FloatArray(ret) { buffer[it] / 32768.0f }
 
                 vad.acceptWaveform(samples)
-                while(!vad.empty()) {vad.pop();}
 
                 val isSpeechDetected = vad.isSpeechDetected()
+                vad.clear()
 
                 runOnUiThread {
                     onVad(isSpeechDetected)
diff --git a/sherpa-onnx/c-api/c-api.cc b/sherpa-onnx/c-api/c-api.cc
index 3f6b35466..1e945ae3e 100644
--- a/sherpa-onnx/c-api/c-api.cc
+++ b/sherpa-onnx/c-api/c-api.cc
@@ -498,6 +498,11 @@ SHERPA_ONNX_API void SherpaOnnxVoiceActivityDetectorPop(
   p->impl->Pop();
 }
 
+SHERPA_ONNX_API void SherpaOnnxVoiceActivityDetectorClear(
+    SherpaOnnxVoiceActivityDetector *p) {
+  p->impl->Clear();
+}
+
 SHERPA_ONNX_API const SherpaOnnxSpeechSegment *
 SherpaOnnxVoiceActivityDetectorFront(SherpaOnnxVoiceActivityDetector *p) {
   const sherpa_onnx::SpeechSegment &segment = p->impl->Front();
diff --git a/sherpa-onnx/c-api/c-api.h b/sherpa-onnx/c-api/c-api.h
index b79b83072..b4b5780a9 100644
--- a/sherpa-onnx/c-api/c-api.h
+++ b/sherpa-onnx/c-api/c-api.h
@@ -580,6 +580,10 @@ SherpaOnnxVoiceActivityDetectorDetected(SherpaOnnxVoiceActivityDetector *p);
 SHERPA_ONNX_API void SherpaOnnxVoiceActivityDetectorPop(
     SherpaOnnxVoiceActivityDetector *p);
 
+// Clear current speech segments.
+SHERPA_ONNX_API void SherpaOnnxVoiceActivityDetectorClear(
+    SherpaOnnxVoiceActivityDetector *p);
+
 // Return the first speech segment.
 // The user has to use SherpaOnnxDestroySpeechSegment() to free the returned
 // pointer to avoid memory leak.
diff --git a/sherpa-onnx/csrc/voice-activity-detector.cc b/sherpa-onnx/csrc/voice-activity-detector.cc
index 05660cd9d..a56caa43e 100644
--- a/sherpa-onnx/csrc/voice-activity-detector.cc
+++ b/sherpa-onnx/csrc/voice-activity-detector.cc
@@ -76,6 +76,8 @@ class VoiceActivityDetector::Impl {
 
   void Pop() { segments_.pop(); }
 
+  void Clear() { std::queue<SpeechSegment>().swap(segments_); }
+
   const SpeechSegment &Front() const { return segments_.front(); }
 
   void Reset() {
diff --git a/sherpa-onnx/csrc/voice-activity-detector.h b/sherpa-onnx/csrc/voice-activity-detector.h
index 61552139b..603bfbe78 100644
--- a/sherpa-onnx/csrc/voice-activity-detector.h
+++ b/sherpa-onnx/csrc/voice-activity-detector.h
@@ -36,6 +36,7 @@ class VoiceActivityDetector {
   void AcceptWaveform(const float *samples, int32_t n);
   bool Empty() const;
   void Pop();
+  void Clear();
   const SpeechSegment &Front() const;
 
   bool IsSpeechDetected() const;
diff --git a/sherpa-onnx/jni/jni.cc b/sherpa-onnx/jni/jni.cc
index f4f0f6d67..92e6e7203 100644
--- a/sherpa-onnx/jni/jni.cc
+++ b/sherpa-onnx/jni/jni.cc
@@ -124,6 +124,8 @@ class SherpaOnnxVad {
 
   void Pop() { vad_.Pop(); }
 
+  void Clear() { vad_.Clear();}
+
   const SpeechSegment &Front() const { return vad_.Front(); }
 
   bool IsSpeechDetected() const { return vad_.IsSpeechDetected(); }
@@ -556,6 +558,14 @@ JNIEXPORT void JNICALL Java_com_k2fsa_sherpa_onnx_Vad_pop(JNIEnv *env,
   model->Pop();
 }
 
+SHERPA_ONNX_EXTERN_C
+JNIEXPORT void JNICALL Java_com_k2fsa_sherpa_onnx_Vad_clear(JNIEnv *env,
+                                                          jobject /*obj*/,
+                                                          jlong ptr) {
+  auto model = reinterpret_cast<sherpa_onnx::SherpaOnnxVad *>(ptr);
+  model->Clear();
+}
+
 // see
 // https://stackoverflow.com/questions/29043872/android-jni-return-multiple-variables
 static jobject NewInteger(JNIEnv *env, int32_t value) {
diff --git a/swift-api-examples/SherpaOnnx.swift b/swift-api-examples/SherpaOnnx.swift
index 53637011c..6c74ba459 100644
--- a/swift-api-examples/SherpaOnnx.swift
+++ b/swift-api-examples/SherpaOnnx.swift
@@ -559,6 +559,10 @@ class SherpaOnnxVoiceActivityDetectorWrapper {
     SherpaOnnxVoiceActivityDetectorPop(vad)
   }
 
+  func clear() {
+    SherpaOnnxVoiceActivityDetectorClear(vad)
+  }
+
   func front() -> SherpaOnnxSpeechSegmentWrapper {
     let p: UnsafePointer<SherpaOnnxSpeechSegment>? = SherpaOnnxVoiceActivityDetectorFront(vad)
     return SherpaOnnxSpeechSegmentWrapper(p: p)
diff --git a/swift-api-examples/generate-subtitles.swift b/swift-api-examples/generate-subtitles.swift
index d06829455..b04b6faef 100644
--- a/swift-api-examples/generate-subtitles.swift
+++ b/swift-api-examples/generate-subtitles.swift
@@ -174,32 +174,31 @@ func run() {
 
   var segments: [SpeechSegment] = []
 
-  while array.count > windowSize {
-    // todo(fangjun): avoid extra copies here
-    vad.acceptWaveform(samples: [Float](array[0..<windowSize]))
-    array = [Float](array[windowSize..<array.count])
-
-    while !vad.isEmpty() {
-      let s = vad.front()
-      vad.pop()
-      let result = recognizer.decode(samples: s.samples)
+  for offset in stride(from: 0, to: array.count, by: windowSize) {
+      let end = min(offset + windowSize, array.count)
+      vad.acceptWaveform(samples: [Float](array[offset ..< end]))
+  }
 
-      segments.append(
-        SpeechSegment(
-          start: Float(s.start) / Float(sampleRate),
-          duration: Float(s.samples.count) / Float(sampleRate),
-          text: result.text))
+  var index: Int = 0
+  while !vad.isEmpty() {
+    let s = vad.front()
+    vad.pop()
+    let result = recognizer.decode(samples: s.samples)
 
-      print(segments.last!)
+    segments.append(
+      SpeechSegment(
+        start: Float(s.start) / Float(sampleRate),
+        duration: Float(s.samples.count) / Float(sampleRate),
+        text: result.text))
 
-    }
+    print(segments.last!)
   }
 
-  let srt = zip(segments.indices, segments).map { (index, element) in
+  let srt: String = zip(segments.indices, segments).map { (index, element) in
     return "\(index+1)\n\(element)"
   }.joined(separator: "\n\n")
 
-  let srtFilename = filePath.stringByDeletingPathExtension + ".srt"
+  let srtFilename: String = filePath.stringByDeletingPathExtension + ".srt"
   do {
     try srt.write(to: srtFilename.fileURL, atomically: true, encoding: .utf8)
   } catch {