Stream speech audio to server as it gets recorded, instead of waiting until end of recording.

content/browser/speech/speech_recognizer.cc

Index: content/browser/speech/speech_recognizer.cc
diff --git a/content/browser/speech/speech_recognizer.cc b/content/browser/speech/speech_recognizer.cc
index edd34195b2ef5600b05ecd612b2d16280d3e79b6..786a6f2e973266b61311e3b05fa9d3aa53dec39e 100644
--- a/content/browser/speech/speech_recognizer.cc
+++ b/content/browser/speech/speech_recognizer.cc
@@ -110,6 +110,7 @@ bool SpeechRecognizer::StartRecording() {
   VLOG(1) << "SpeechRecognizer starting record.";
   num_samples_recorded_ = 0;
   audio_controller_->Record();
+  previous_audio_chunk_.clear();
 
   return true;
 }
@@ -141,34 +142,33 @@ void SpeechRecognizer::StopRecording() {
   VLOG(1) << "SpeechRecognizer stopping record.";
   audio_controller_->Close();
   audio_controller_ = NULL;  // Releases the ref ptr.
-  encoder_->Flush();
 
   delegate_->DidCompleteRecording(caller_id_);
 
-  // Since the http request takes a single string as POST data, allocate
-  // one and copy over bytes from the audio buffers to the string.
-  // And If we haven't got any audio yet end the recognition sequence here.
-  string mime_type = encoder_->mime_type();
-  string data;
-  encoder_->GetEncodedData(&data);
+  // Get any last bits of encoded data left.
+  encoder_->Flush();
+  string encoded_data;
+  encoder_->GetEncodedDataAndClear(&encoded_data);
   encoder_.reset();
 
-  if (data.empty()) {
+  // If we haven't got any audio yet end the recognition sequence here.
+  if (request_ == NULL) {
     // Guard against the delegate freeing us until we finish our job.
     scoped_refptr<SpeechRecognizer> me(this);
     delegate_->DidCompleteRecognition(caller_id_);
   } else {
-    DCHECK(!request_.get());
-    request_.reset(new SpeechRecognitionRequest(
-        Profile::GetDefaultRequestContext(), this));
-    request_->Send(language_, grammar_, hardware_info_, origin_url_,
-                   mime_type, data);
+    // UploadAudioChunk requires a non-empty buffer. So we check if there was
+    // any data available since last time we sent and if not try to send the
+    // last chunk again (i.e. repeat the last 100ms of audio). If nothing was
+    // recorded yet, we just send a whitespace string.

[bulach, 2011/03/03 20:05:00]

hmm, I haven't followed the other CL this one depends, but wouldn't it be
simpler to have just a "FinishChunk" (or something similar)? worst case, it may
be simpler to move the "emptiness" test down to UploadAudioChunk rather than
here?

also, why is it necessary to send previoud_audio_chunk again?

if both we can't work around these limitations, please also clarify why repeat
the last chunk.

+    if (encoded_data.empty()) {
+      encoded_data = !previous_audio_chunk_.empty() ?
+                     previous_audio_chunk_ : " ";
+    }
+    request_->UploadAudioChunk(encoded_data, true);
   }
 }
 
-void SpeechRecognizer::ReleaseAudioBuffers() {
-}
-
 // Invoked in the audio thread.
 void SpeechRecognizer::OnError(AudioInputController* controller,
                                int error_code) {
@@ -214,7 +214,6 @@ void SpeechRecognizer::HandleOnData(string* data) {
   const short* samples = reinterpret_cast<const short*>(data->data());
   DCHECK((data->length() % sizeof(short)) == 0);
   int num_samples = data->length() / sizeof(short);
-
   encoder_->Encode(samples, num_samples);
   float rms;
   endpointer_.ProcessAudio(samples, num_samples, &rms);
@@ -222,6 +221,21 @@ void SpeechRecognizer::HandleOnData(string* data) {
   delete data;
   num_samples_recorded_ += num_samples;
 
+  if (request_ == NULL) {
+    // This was the first audio packet recorded, so start a request to the
+    // server to send the data.
+    request_.reset(new SpeechRecognitionRequest(
+        Profile::GetDefaultRequestContext(), this));
+    request_->Start(language_, grammar_, hardware_info_, origin_url_,
+                    encoder_->mime_type());
+  }
+
+  string encoded_data;
+  encoder_->GetEncodedDataAndClear(&encoded_data);
+  DCHECK(!encoded_data.empty());
+  request_->UploadAudioChunk(encoded_data, false);
+  previous_audio_chunk_ = encoded_data;
+
   if (endpointer_.IsEstimatingEnvironment()) {
     // Check if we have gathered enough audio for the endpointer to do
     // environment estimation and should move on to detect speech/end of speech.