Speech refactoring: Reimplemented SpeechRecognitionManagerImpl as a FSM. (CL1.7)

content/browser/speech/speech_recognition_manager_impl.cc

 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "content/browser/speech/speech_recognition_manager_impl.h"
 
 #include "base/bind.h"
+#include "base/memory/singleton.h"
 #include "content/browser/browser_main_loop.h"
-#include "content/browser/renderer_host/render_view_host_impl.h"
-#include "content/browser/speech/input_tag_speech_dispatcher_host.h"
+#include "content/browser/speech/google_one_shot_remote_engine.h"
+#include "content/browser/speech/speech_recognition_engine.h"
+#include "content/browser/speech/speech_recognizer_impl.h"
 #include "content/public/browser/browser_thread.h"
 #include "content/public/browser/content_browser_client.h"
-#include "content/public/browser/speech_recognizer.h"
-#include "content/public/browser/render_view_host_delegate.h"
 #include "content/public/browser/resource_context.h"
+#include "content/public/browser/speech_recognition_event_listener.h"
 #include "content/public/browser/speech_recognition_manager_delegate.h"
-#include "content/public/browser/speech_recognition_preferences.h"
-#include "content/public/common/view_type.h"
+#include "content/public/browser/speech_recognition_session_config.h"
+#include "content/public/browser/speech_recognition_session_context.h"
+#include "content/public/common/speech_recognition_result.h"
 #include "media/audio/audio_manager.h"
 
+using base::Callback;
+using base::Unretained;
 using content::BrowserMainLoop;
 using content::BrowserThread;
-using content::RenderViewHostImpl;
+using content::SpeechRecognitionError;
+using content::SpeechRecognitionEventListener;
 using content::SpeechRecognitionManager;
-using content::SpeechRecognitionManagerDelegate;
+using content::SpeechRecognitionResult;
+using content::SpeechRecognitionSessionContext;
+using content::SpeechRecognitionSessionConfig;
+
+namespace {
+
+// A dummy implementation of the SpeechRecognitionManagerDelegate interface
+// used when no delegate has been passed to the SpeechRecognitionManagerImpl.
+class VoidRecognitionManagerDelegate :
+    public content::SpeechRecognitionManagerDelegate {
+ public:
+  static VoidRecognitionManagerDelegate* GetInstance() {
+    return Singleton<VoidRecognitionManagerDelegate>::get();
+  }
+  virtual void GetDiagnosticInformation(
+      bool* can_report_metrics, std::string* request_info) OVERRIDE {}
+  virtual bool IsRecognitionAllowed(int session_id) OVERRIDE { return false; }
+  virtual void ShowRecognitionRequested(int session_id) OVERRIDE {}
+  virtual void ShowWarmUp(int session_id) OVERRIDE {}
+  virtual void ShowRecognizing(int session_id) OVERRIDE {}
+  virtual void ShowRecording(int session_id) OVERRIDE {}
+  virtual void ShowInputVolume(
+      int session_id, float volume, float noise_volume) OVERRIDE {}
+  virtual void ShowError(int session_id,
+      const content::SpeechRecognitionError& error) OVERRIDE {}
+  virtual void DoClose(int session_id) OVERRIDE {}
+
+ private:
+  VoidRecognitionManagerDelegate() {}
+  virtual ~VoidRecognitionManagerDelegate() {}
+  friend struct DefaultSingletonTraits<VoidRecognitionManagerDelegate>;
+};
+
+}  // namespace
+
+namespace content {
+const int SpeechRecognitionManager::kSessionIDInvalid = 0;
 
 SpeechRecognitionManager* SpeechRecognitionManager::GetInstance() {
   return speech::SpeechRecognitionManagerImpl::GetInstance();
 }
+}  // namespace content
 
 namespace speech {
 
-struct SpeechRecognitionManagerImpl::SpeechRecognitionParams {
-  SpeechRecognitionParams(
-      InputTagSpeechDispatcherHost* delegate,
-      int session_id,
-      int render_process_id,
-      int render_view_id,
-      const gfx::Rect& element_rect,
-      const std::string& language,
-      const std::string& grammar,
-      const std::string& origin_url,
-      net::URLRequestContextGetter* context_getter,
-      content::SpeechRecognitionPreferences* recognition_prefs)
-      : delegate(delegate),
-        session_id(session_id),
-        render_process_id(render_process_id),
-        render_view_id(render_view_id),
-        element_rect(element_rect),
-        language(language),
-        grammar(grammar),
-        origin_url(origin_url),
-        context_getter(context_getter),
-        recognition_prefs(recognition_prefs) {
-  }
-
-  InputTagSpeechDispatcherHost* delegate;
-  int session_id;
-  int render_process_id;
-  int render_view_id;
-  gfx::Rect element_rect;
-  std::string language;
-  std::string grammar;
-  std::string origin_url;
-  net::URLRequestContextGetter* context_getter;
-  content::SpeechRecognitionPreferences* recognition_prefs;
-};
-
 SpeechRecognitionManagerImpl* SpeechRecognitionManagerImpl::GetInstance() {
   return Singleton<SpeechRecognitionManagerImpl>::get();
 }
 
 SpeechRecognitionManagerImpl::SpeechRecognitionManagerImpl()
-    : can_report_metrics_(false),
-      recording_session_id_(0) {
-  delegate_.reset(content::GetContentClient()->browser()->
-                  GetSpeechRecognitionManagerDelegate());
+    : interactive_session_id_(kSessionIDInvalid),
+      last_session_id_(kSessionIDInvalid),
+      is_dispatching_event_(false) {
+  delegate_ = content::GetContentClient()->browser()->
+      GetSpeechRecognitionManagerDelegate();
+  // In lack of one being provided, instantiate a void delegate so we can avoid
+  // unaesthetic "if (delegate_ != NULL)" statements.
+  if (delegate_ == NULL)
+    delegate_ = VoidRecognitionManagerDelegate::GetInstance();
 }
 
 SpeechRecognitionManagerImpl::~SpeechRecognitionManagerImpl() {
-  while (requests_.begin() != requests_.end())
-    CancelRecognition(requests_.begin()->first);
-}
+  // Recognition sessions will be aborted by the corresponding destructors.
+  sessions_.clear();
+}
+
+int SpeechRecognitionManagerImpl::CreateSession(
+    const SpeechRecognitionSessionConfig& config,
+    SpeechRecognitionEventListener* event_listener) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+
+  const int session_id = GetNextSessionID();
+  DCHECK(!SessionExists(session_id));
+  // Set-up the new session.
+  Session& session = sessions_[session_id];
+  session.id = session_id;
+  session.event_listener = event_listener;
+  session.context = config.initial_context;
+
+  // TODO(primiano) Is this check enough just on creation or shall we move/copy
+  // it on SessionStart in order to repeat the check every time?.
+  if (!delegate_->IsRecognitionAllowed(session_id)) {
+    sessions_.erase(session_id);
+    return kSessionIDInvalid;
+  }
+
+  std::string hardware_info;
+  bool can_report_metrics;
+  delegate_->GetDiagnosticInformation(&can_report_metrics, &hardware_info);
+
+  GoogleOneShotRemoteEngineConfig remote_engine_config;
+  remote_engine_config.language = config.language;
+  remote_engine_config.grammar = config.grammar;
+  remote_engine_config.audio_sample_rate =
+      SpeechRecognizerImpl::kAudioSampleRate;
+  remote_engine_config.audio_num_bits_per_sample =
+     SpeechRecognizerImpl::kNumBitsPerAudioSample;
+  remote_engine_config.filter_profanities = config.filter_profanities;
+  remote_engine_config.hardware_info = hardware_info;
+  remote_engine_config.origin_url = can_report_metrics ? config.origin_url : "";
+
+  GoogleOneShotRemoteEngine* google_remote_engine =
+      new GoogleOneShotRemoteEngine(config.url_request_context_getter);
+  google_remote_engine->SetConfig(remote_engine_config);
+
+  session.recognizer = new SpeechRecognizerImpl(this,
+                                                session_id,
+                                                google_remote_engine);
+  return session_id;
+}
+
+void SpeechRecognitionManagerImpl::StartSession(int session_id) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  DCHECK(SessionExists(session_id));
+
+  // If there is another interactive session, detach prior to start the new one.
+  if (interactive_session_id_ != kSessionIDInvalid &&
+      interactive_session_id_ != session_id) {
+     DetachSession(interactive_session_id_);
+  }
+
+   BrowserThread::PostTask(BrowserThread::IO, FROM_HERE,
+     base::Bind(&SpeechRecognitionManagerImpl::DispatchEvent, Unretained(this),
+                session_id, FSMEventArgs(EVENT_START)));
+}
+
+void SpeechRecognitionManagerImpl::AbortSession(int session_id) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  DCHECK(SessionExists(session_id));
+
+  BrowserThread::PostTask(BrowserThread::IO, FROM_HERE,
+      base::Bind(&SpeechRecognitionManagerImpl::DispatchEvent, Unretained(this),
+                 session_id, FSMEventArgs(EVENT_ABORT)));
+}
+
+void SpeechRecognitionManagerImpl::StopAudioCaptureForSession(int session_id) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  DCHECK(SessionExists(session_id));
+
+  BrowserThread::PostTask(BrowserThread::IO, FROM_HERE,
+      base::Bind(&SpeechRecognitionManagerImpl::DispatchEvent, Unretained(this),
+                 session_id, FSMEventArgs(EVENT_STOP_CAPTURE)));
+}
+
+void SpeechRecognitionManagerImpl::DetachSession(int session_id) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  DCHECK(SessionExists(session_id));
+
+  BrowserThread::PostTask(BrowserThread::IO, FROM_HERE,
+      base::Bind(&SpeechRecognitionManagerImpl::DispatchEvent, Unretained(this),
+                 session_id, FSMEventArgs(EVENT_DETACH)));
+}
+
+// Here begins the SpeechRecognitionEventListener interface implementation,
+// which will simply relay the events to the proper listener registered for the
+// particular session (most likely InputTagSpeechDispatcherHost) and intercept
+// some of them to provide UI notifications.
+
+void SpeechRecognitionManagerImpl::OnRecognitionStart(int session_id) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  if (!SessionExists(session_id))
+    return;
+
+  DCHECK_EQ(interactive_session_id_, session_id);
+  delegate_->ShowWarmUp(session_id);
+  GetListener(session_id)->OnRecognitionStart(session_id);
+}
+
+void SpeechRecognitionManagerImpl::OnAudioStart(int session_id) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  if (!SessionExists(session_id))
+    return;
+
+  DCHECK_EQ(interactive_session_id_, session_id);
+  delegate_->ShowRecording(session_id);
+  GetListener(session_id)->OnAudioStart(session_id);
+}
+
+void SpeechRecognitionManagerImpl::OnEnvironmentEstimationComplete(
+    int session_id) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  if (!SessionExists(session_id))
+    return;
+
+  DCHECK_EQ(interactive_session_id_, session_id);
+  GetListener(session_id)->OnEnvironmentEstimationComplete(session_id);
+}
+
+void SpeechRecognitionManagerImpl::OnSoundStart(int session_id) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  if (!SessionExists(session_id))
+    return;
+
+  DCHECK_EQ(interactive_session_id_, session_id);
+  GetListener(session_id)->OnSoundStart(session_id);
+}
+
+void SpeechRecognitionManagerImpl::OnSoundEnd(int session_id) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  if (!SessionExists(session_id))
+    return;
+
+  GetListener(session_id)->OnSoundEnd(session_id);
+}
+
+void SpeechRecognitionManagerImpl::OnAudioEnd(int session_id) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  if (!SessionExists(session_id))
+    return;
+
+  // OnAudioEnd can also be raised after an abort request, when the session is
+  // not interactive anymore.
+  if (interactive_session_id_ == session_id)
+    delegate_->ShowRecognizing(session_id);
+
+  GetListener(session_id)->OnAudioEnd(session_id);
+}
+
+void SpeechRecognitionManagerImpl::OnRecognitionResult(
+    int session_id, const content::SpeechRecognitionResult& result) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  if (!SessionExists(session_id))
+    return;
+
+  GetListener(session_id)->OnRecognitionResult(session_id, result);
+  FSMEventArgs event_args(EVENT_RECOGNITION_RESULT);
+  BrowserThread::PostTask(BrowserThread::IO, FROM_HERE,
+    base::Bind(&SpeechRecognitionManagerImpl::DispatchEvent, Unretained(this),
+               session_id, event_args));
+}
+
+void SpeechRecognitionManagerImpl::OnRecognitionError(
+    int session_id, const content::SpeechRecognitionError& error) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  if (!SessionExists(session_id))
+    return;
+
+  GetListener(session_id)->OnRecognitionError(session_id, error);
+  FSMEventArgs event_args(EVENT_RECOGNITION_ERROR);
+  event_args.speech_error = error;
+  BrowserThread::PostTask(BrowserThread::IO, FROM_HERE,
+    base::Bind(&SpeechRecognitionManagerImpl::DispatchEvent, Unretained(this),
+               session_id, event_args));
+}
+
+void SpeechRecognitionManagerImpl::OnAudioLevelsChange(
+    int session_id, float volume, float noise_volume) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  if (!SessionExists(session_id))
+    return;
+
+  delegate_->ShowInputVolume(session_id, volume, noise_volume);
+  GetListener(session_id)->OnAudioLevelsChange(session_id, volume,
+                                               noise_volume);
+}
+
+void SpeechRecognitionManagerImpl::OnRecognitionEnd(int session_id) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  if (!SessionExists(session_id))
+    return;
+
+  GetListener(session_id)->OnRecognitionEnd(session_id);
+  BrowserThread::PostTask(BrowserThread::IO, FROM_HERE,
+    base::Bind(&SpeechRecognitionManagerImpl::DispatchEvent, Unretained(this),
+               session_id, FSMEventArgs(EVENT_RECOGNITION_ENDED)));
+}
+
+// TODO(primiano) After CL2: if we see that both InputTagDispatcherHost and
+// SpeechRecognitionDispatcherHost do the same lookup operations, implement the
+// lookup method directly here.
+int SpeechRecognitionManagerImpl::LookupSessionByContext(
+    Callback<bool(const SpeechRecognitionSessionContext&)> matcher) const {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  SessionsTable::const_iterator iter;
+  // Note: the callback (matcher) must NEVER perform non-const calls on us.
+  for(iter = sessions_.begin(); iter != sessions_.end(); iter++) {
+    const int session_id = iter->first;
+    const Session& session = iter->second;
+    bool matches = matcher.Run(session.context);
+    if (matches)
+      return session_id;
+  }
+  return kSessionIDInvalid;
+}
+
+SpeechRecognitionSessionContext
+SpeechRecognitionManagerImpl::GetSessionContext(int session_id) const {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  SessionsTable::const_iterator iter = sessions_.find(session_id);
+  DCHECK(iter != sessions_.end());
+  return const_cast<SpeechRecognitionSessionContext&>(iter->second.context);
+}
+
+void SpeechRecognitionManagerImpl::AbortAllSessionsForListener(
+    SpeechRecognitionEventListener* listener) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  SessionsTable::iterator it = sessions_.begin();
+  // AbortSession is asynchronous and the session will not be removed from the
+  // collection while we are iterating over it.
+  while (it != sessions_.end()) {
+    if (it->second.event_listener == listener)
+      AbortSession(it->first);
+    ++it;
+  }
+}
+
+// -----------------------  Core FSM implementation ---------------------------
+void SpeechRecognitionManagerImpl::DispatchEvent(int session_id,
+                                                 FSMEventArgs event_args) {
+  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
+  if (!SessionExists(session_id))
+    return;
+
+  Session& session = sessions_[session_id];
+  DCHECK_LE(session.state, STATE_MAX_VALUE);
+  DCHECK_LE(event_args.event, EVENT_MAX_VALUE);
+
+  // Event dispatching must be sequential, otherwise it will break all the rules
+  // and the assumptions of the finite state automata model.
+  DCHECK(!is_dispatching_event_);
+  is_dispatching_event_ = true;
+
+  // Pedantic preconditions consistency checks.
+  if (session.state == STATE_INTERACTIVE)
+    DCHECK_EQ(interactive_session_id_, session_id);
+
+  if (session.state == STATE_DETACHED ||
+      session.state == STATE_WAITING_FOR_DELETION) {
+    DCHECK_NE(interactive_session_id_, session_id);
+  }
+
+  session.state = ExecuteTransitionAndGetNextState(session, event_args);
+
+  is_dispatching_event_ = false;
+}
+
+// This FSM handles the evolution of each session, from the viewpoint of the
+// interaction with the user (that may be either the browser end-user which
+// interacts with UI bubbles, or JS developer intracting with JS methods).
+// All the events received by the SpeechRecognizerImpl instances (one for each
+// session) are always routed to the SpeechRecognitionEventListener(s)
+// regardless the choices taken in this FSM.
+SpeechRecognitionManagerImpl::FSMState
+SpeechRecognitionManagerImpl::ExecuteTransitionAndGetNextState(
+    Session& session, const FSMEventArgs& event_args) {
+  // Some notes for the code below:
+  // - A session can be deleted only if it is not active, thus only if it ended
+  //   spontaneously or we issued a prior SessionAbort. In these cases, we must
+  //   wait for a RECOGNITION_ENDED event (which is guaranteed to come always at
+  //   last by the SpeechRecognizer) in order to free resources gracefully.
+  // - Use SessionDelete only when absolutely sure that the recognizer is not
+  //   active. Prefer SessionAbort, which will do it gracefully, otherwise.
+  const FSMEvent event = event_args.event;
+  switch (session.state) {
+    case STATE_IDLE:
+      // Session has just been created or had an error while interactive (thus,
+      // it is still interactive).
+      switch (event) {
+        case EVENT_START:
+          return SessionStart(session, event_args);
+        case EVENT_ABORT:
+        case EVENT_DETACH:
+          return SessionAbort(session, event_args);
+        case EVENT_STOP_CAPTURE:
+        case EVENT_RECOGNITION_ENDED:
+          return DoNothing(session, event_args);

[Satish, 2012/04/23 10:25:17]

why are these 2 events valid in this state? shouldn't they not be feasible?

[Primiano Tucci (use gerrit), 2012/04/23 12:52:25]

STOP_CAPTURE is feasible (though it makes no sense) since it is an event
requested through a public method. According to the contract of SRM (see
content/public/browser/speech_recognition_manager.h) "the SRM has the further
responsibility of handling separately and reliably (taking into account also
call sequences that might not make sense, e.g., two subsequent AbortSession
calls)".
From a more practical viewpoint, it is not expected to happen in the current CL.
But it might happen in the near future when we'll have JS support. In that case
we can end up in this case if the StopAudioCapture is called on the JS object
without prior starting recognition.

EVENT_RECOGNITION_ENDED is feasible (and happens) in the case of an error.
Perhaps the name "IDLE" is not 99.9% correct, since there is the current corner
case in which we are not exactly IDLE, yet for a very short time interval: SRM
is in the state INTERACTIVE and receives a RECOGNITION_ERROR event -> the bubble
error is shown (through the delegate) calling SessionReportError and SRM
transits to IDLE state.  However, since the OnError event is always guaranteed
to precede the OnRecognitionEnd event, the RECOGNITION_ENDED event will be
received after the transition to the IDLE state.
This is a consequence of merging together the previous states IDLE and
ENDED_WITH_ERROR (which were almost identical, expect for this detail). Any
comment is welcome, and if you prefer, we can switch back adding a separate
ENDED_WITH_ERROR state.

[Satish, 2012/04/24 08:51:32]

Makes sense. Can you mention a shorter version of this as a comment here?

+        case EVENT_RECOGNITION_RESULT:
+        case EVENT_RECOGNITION_ERROR:
+          return NotFeasible(session, event_args);
+      }
+      break;
+    case STATE_INTERACTIVE:
+      // The recognizer can be either capturing audio or waiting for a result.
+      switch (event) {
+        case EVENT_RECOGNITION_RESULT:
+          // TODO(primiano) Valid only in single shot mode. Review in next CLs.
+          return SessionDetach(session, event_args);
+        case EVENT_DETACH:
+          return SessionAbortIfCapturingAudioOrDetach(session, event_args);
+        case EVENT_STOP_CAPTURE:
+          return SessionStopAudioCapture(session, event_args);
+        case EVENT_ABORT:
+          return SessionAbort(session, event_args);
+        case EVENT_RECOGNITION_ERROR:
+          return SessionReportError(session, event_args);
+        case EVENT_RECOGNITION_ENDED:
+          // If we're still interactive it means that no result was received
+          // in the meanwhile (otherwise we'd have been detached).
+          return SessionReportNoMatch(session, event_args);
+        case EVENT_START:
+          return DoNothing(session, event_args);
+      }
+      break;
+    case STATE_DETACHED:

[Satish, 2012/04/23 10:25:17]

As I mentioned earlier the concept of a 'detached' session isn't clear still
while the others are quite clear. STATE_DETACHED and STATE_WAITING_FOR_DELETION
seem very close in purpose, both can't get the session back into active state
and are just different stages of cleaning up the session. Can we combine them
into a single state?

[Primiano Tucci (use gerrit), 2012/04/23 12:52:25]

Ah sorry, I thought the earlier comment was just referred to DETACHABLE <>
DETACHED states.
My idea that DETACHED (we can change the name if we find a more suitable one,
perhaps BACKGROUND?) shall be used most in the near future for continuous
recognition sessions that don't use the bubble (or any other UI interaction for
start/stop), that are, upcoming JS APIs and speech input extensions (that
currently are "managing" speech on their own). Merging WAITING_FOR_DELETION and
DETACHED should work in the current implementation, but perhaps (I am not 100%
sure about this now) DETACHED will be changed in the near future for continuous
recognition (so the merge will probably need to be reverted).
Should I add a TODO to see if we can merge them in future, after the continuous
speech APIs??

[Satish, 2012/04/24 08:51:32]

Yes renaming DETACH to BACKGROUND would be clearer and goes well with
INTERACTIVE as well. No need to merge otherwise.

+      switch (event) {
+        case EVENT_ABORT:
+          return SessionAbort(session, event_args);
+        case EVENT_RECOGNITION_ENDED:
+          return SessionDelete(session, event_args);
+        case EVENT_START:
+        case EVENT_STOP_CAPTURE:
+        case EVENT_RECOGNITION_RESULT:
+        case EVENT_RECOGNITION_ERROR:
+          return DoNothing(session, event_args);
+        case EVENT_DETACH:
+          return NotFeasible(session, event_args);
+      }
+      break;
+    case STATE_WAITING_FOR_DELETION:
+      switch (event) {
+        case EVENT_RECOGNITION_ENDED:
+          return SessionDelete(session, event_args);
+        case EVENT_ABORT:
+        case EVENT_START:
+        case EVENT_STOP_CAPTURE:
+        case EVENT_DETACH:
+        case EVENT_RECOGNITION_RESULT:
+        case EVENT_RECOGNITION_ERROR:
+          return DoNothing(session, event_args);
+      }
+      break;
+  }
+  return NotFeasible(session, event_args);
+}
+
+// ----------- Contract for all the FSM evolution functions below -------------
+//  - Are guaranteed to be executed in the IO thread;
+//  - Are guaranteed to be not reentrant (themselves and each other);
+//  - event_args members are guaranteed to be stable during the call;
+
+SpeechRecognitionManagerImpl::FSMState
+SpeechRecognitionManagerImpl::SessionStart(Session& session,
+                                           const FSMEventArgs& event_args) {
+ if (interactive_session_id_ != 0)
+    delegate_->DoClose(interactive_session_id_);
+  interactive_session_id_ = session.id;
+  delegate_->ShowRecognitionRequested(session.id);
+  session.recognizer->StartRecognition();
+  return STATE_INTERACTIVE;
+}
+
+SpeechRecognitionManagerImpl::FSMState
+SpeechRecognitionManagerImpl::SessionAbort(Session& session,
+                                           const FSMEventArgs& event_args) {
+  if (interactive_session_id_ == session.id) {
+    interactive_session_id_ = kSessionIDInvalid;
+    delegate_->DoClose(session.id);
+  }
+
+  // If abort was requested while the recognizer was inactive, delete directly.
+  if (session.recognizer == NULL || !session.recognizer->IsActive())
+    return SessionDelete(session, event_args);
+
+  // Otherwise issue an abort and delete gracefully, waiting for a
+  // RECOGNITION_ENDED event first.
+  session.recognizer->AbortRecognition();
+  return STATE_WAITING_FOR_DELETION;
+}
+
+SpeechRecognitionManagerImpl::FSMState
+SpeechRecognitionManagerImpl::SessionStopAudioCapture(
+    Session& session, const FSMEventArgs& event_args) {
+  DCHECK(session.recognizer != NULL);
+  DCHECK(session.recognizer->IsActive());
+  if (session.recognizer->IsCapturingAudio())
+    session.recognizer->StopAudioCapture();
+  return STATE_INTERACTIVE;
+}
+
+SpeechRecognitionManagerImpl::FSMState
+SpeechRecognitionManagerImpl::SessionAbortIfCapturingAudioOrDetach(
+    Session& session, const FSMEventArgs& event_args) {
+  DCHECK_EQ(interactive_session_id_, session.id);
+
+  DCHECK(session.recognizer != NULL);
+  DCHECK(session.recognizer->IsActive());
+  if (session.recognizer->IsCapturingAudio())
+    return SessionAbort(session, event_args);
+
+  interactive_session_id_ = kSessionIDInvalid;
+  delegate_->DoClose(session.id);
+  return STATE_DETACHED;
+}
+
+
+SpeechRecognitionManagerImpl::FSMState
+SpeechRecognitionManagerImpl::SessionDetach(Session& session,
+                                            const FSMEventArgs& event_args) {
+  DCHECK_EQ(interactive_session_id_, session.id);
+  interactive_session_id_ = kSessionIDInvalid;
+  delegate_->DoClose(session.id);
+  return STATE_DETACHED;
+}
+
+SpeechRecognitionManagerImpl::FSMState
+SpeechRecognitionManagerImpl::SessionReportError(
+    Session& session, const FSMEventArgs& event_args) {
+  DCHECK_EQ(interactive_session_id_, session.id);
+  delegate_->ShowError(session.id, event_args.speech_error);
+  return STATE_IDLE;
+}
+
+SpeechRecognitionManagerImpl::FSMState
+SpeechRecognitionManagerImpl::SessionReportNoMatch(
+    Session& session, const FSMEventArgs& event_args) {
+  DCHECK_EQ(interactive_session_id_, session.id);
+  delegate_->ShowError(
+      session.id,
+      SpeechRecognitionError(content::SPEECH_RECOGNITION_ERROR_NO_MATCH));
+  return STATE_IDLE;
+}
+
+SpeechRecognitionManagerImpl::FSMState
+SpeechRecognitionManagerImpl::SessionDelete(Session& session,
+                                            const FSMEventArgs& event_args) {
+  DCHECK(session.recognizer == NULL || !session.recognizer->IsActive());
+  if (interactive_session_id_ == session.id) {
+    interactive_session_id_ = kSessionIDInvalid;
+    delegate_->DoClose(session.id);
+  }
+  sessions_.erase(session.id);
+  // Next state is ininfluent, the session will be deleted afterwards.
+  return STATE_WAITING_FOR_DELETION;
+}
+
+SpeechRecognitionManagerImpl::FSMState
+SpeechRecognitionManagerImpl::DoNothing(Session& session,
+                                        const FSMEventArgs& event_args) {
+  return session.state;
+}
+
+SpeechRecognitionManagerImpl::FSMState
+SpeechRecognitionManagerImpl::NotFeasible(Session& session,
+                                          const FSMEventArgs& event_args) {
+  NOTREACHED() << "Unfeasible event " << event_args.event
+               << " in state " << session.state
+               << " for session " << session.id;
+  return session.state;
+}
+
+int SpeechRecognitionManagerImpl::GetNextSessionID() {
+  ++last_session_id_;
+  // Deal with wrapping of last_session_id_. (How civilized).
+  if (last_session_id_ <= 0)
+    last_session_id_ = 1;
+  return last_session_id_;
+}
+
+bool SpeechRecognitionManagerImpl::SessionExists(int session_id) const {
+  return sessions_.find(session_id) != sessions_.end();
+}
+
+SpeechRecognitionEventListener* SpeechRecognitionManagerImpl::GetListener(
+    int session_id) const {
+  return sessions_.find(session_id)->second.event_listener;
+}
+
 
 bool SpeechRecognitionManagerImpl::HasAudioInputDevices() {
   return BrowserMainLoop::GetAudioManager()->HasAudioInputDevices();
 }
 
 bool SpeechRecognitionManagerImpl::IsCapturingAudio() {
   return BrowserMainLoop::GetAudioManager()->IsRecordingInProcess();
 }
 
 string16 SpeechRecognitionManagerImpl::GetAudioInputDeviceModel() {
   return BrowserMainLoop::GetAudioManager()->GetAudioInputDeviceModel();
 }
 
-bool SpeechRecognitionManagerImpl::HasPendingRequest(int session_id) const {
-  return requests_.find(session_id) != requests_.end();
-}
-
-InputTagSpeechDispatcherHost* SpeechRecognitionManagerImpl::GetDelegate(
-    int session_id) const {
-  return requests_.find(session_id)->second.delegate;
-}
-
 void SpeechRecognitionManagerImpl::ShowAudioInputSettings() {
   // Since AudioManager::ShowAudioInputSettings can potentially launch external
   // processes, do that in the FILE thread to not block the calling threads.
   if (!BrowserThread::CurrentlyOn(BrowserThread::FILE)) {
     BrowserThread::PostTask(
         BrowserThread::FILE, FROM_HERE,
         base::Bind(&SpeechRecognitionManagerImpl::ShowAudioInputSettings,
                    base::Unretained(this)));
     return;
   }
 
   media::AudioManager* audio_manager = BrowserMainLoop::GetAudioManager();
   DCHECK(audio_manager->CanShowAudioInputSettings());
   if (audio_manager->CanShowAudioInputSettings())
     audio_manager->ShowAudioInputSettings();
 }
 
-void SpeechRecognitionManagerImpl::StartRecognition(
-    InputTagSpeechDispatcherHost* delegate,
-    int session_id,
-    int render_process_id,
-    int render_view_id,
-    const gfx::Rect& element_rect,
-    const std::string& language,
-    const std::string& grammar,
-    const std::string& origin_url,
-    net::URLRequestContextGetter* context_getter,
-    content::SpeechRecognitionPreferences* recognition_prefs) {
-  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
-  BrowserThread::PostTask(
-      BrowserThread::UI, FROM_HERE,
-      base::Bind(
-          &SpeechRecognitionManagerImpl::CheckRenderViewTypeAndStartRecognition,
-          base::Unretained(this),
-          SpeechRecognitionParams(
-              delegate, session_id, render_process_id, render_view_id,
-              element_rect, language, grammar, origin_url, context_getter,
-              recognition_prefs)));
-}
-
-void SpeechRecognitionManagerImpl::CheckRenderViewTypeAndStartRecognition(
-    const SpeechRecognitionParams& params) {
-  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::UI));
-
-  RenderViewHostImpl* render_view_host = RenderViewHostImpl::FromID(
-      params.render_process_id, params.render_view_id);
-  if (!render_view_host || !render_view_host->GetDelegate())
-    return;
-
-  // For host delegates other than VIEW_TYPE_WEB_CONTENTS we can't reliably show
-  // a popup, including the speech input bubble. In these cases for privacy
-  // reasons we don't want to start recording if the user can't be properly
-  // notified. An example of this is trying to show the speech input bubble
-  // within an extension popup: http://crbug.com/92083. In these situations the
-  // speech input extension API should be used instead.
-  if (render_view_host->GetDelegate()->GetRenderViewType() ==
-      content::VIEW_TYPE_WEB_CONTENTS) {
-    BrowserThread::PostTask(
-        BrowserThread::IO, FROM_HERE,
-        base::Bind(&SpeechRecognitionManagerImpl::ProceedStartingRecognition,
-        base::Unretained(this), params));
-  }
-}
-
-void SpeechRecognitionManagerImpl::ProceedStartingRecognition(
-    const SpeechRecognitionParams& params) {
-  DCHECK(BrowserThread::CurrentlyOn(BrowserThread::IO));
-  DCHECK(!HasPendingRequest(params.session_id));
-
-  if (delegate_.get()) {
-    delegate_->ShowRecognitionRequested(
-        params.session_id, params.render_process_id, params.render_view_id,
-        params.element_rect);
-    delegate_->GetRequestInfo(&can_report_metrics_, &request_info_);
-  }
-
-  Request* request = &requests_[params.session_id];
-  request->delegate = params.delegate;
-  request->recognizer = content::SpeechRecognizer::Create(
-      this, params.session_id, params.language, params.grammar,
-      params.context_getter, params.recognition_prefs->FilterProfanities(),
-      request_info_, can_report_metrics_ ? params.origin_url : "");
-  request->is_active = false;
-
-  StartRecognitionForRequest(params.session_id);
-}
-
-void SpeechRecognitionManagerImpl::StartRecognitionForRequest(int session_id) {
-  SpeechRecognizerMap::iterator request = requests_.find(session_id);
-  if (request == requests_.end()) {
-    NOTREACHED();
-    return;
-  }
-
-  // We should not currently be recording for the session.
-  CHECK(recording_session_id_ != session_id);
-
-  // If we are currently recording audio for another session, abort it cleanly.
-  if (recording_session_id_)
-    CancelRecognitionAndInformDelegate(recording_session_id_);
-  recording_session_id_ = session_id;
-  requests_[session_id].is_active = true;
-  requests_[session_id].recognizer->StartRecognition();
-  if (delegate_.get())
-    delegate_->ShowWarmUp(session_id);
-}
-
-void SpeechRecognitionManagerImpl::CancelRecognitionForRequest(int session_id) {
-  // Ignore if the session id was not in our active recognizers list because the
-  // user might have clicked more than once, or recognition could have been
-  // ended due to other reasons before the user click was processed.
-  if (!HasPendingRequest(session_id))
-    return;
-
-  CancelRecognitionAndInformDelegate(session_id);
-}
-
-void SpeechRecognitionManagerImpl::FocusLostForRequest(int session_id) {
-  // See above comment.
-  if (!HasPendingRequest(session_id))
-    return;
-
-  // If this is an ongoing recording or if we were displaying an error message
-  // to the user, abort it since user has switched focus. Otherwise
-  // recognition has started and keep that going so user can start speaking to
-  // another element while this gets the results in parallel.
-  if (recording_session_id_ == session_id || !requests_[session_id].is_active)
-    CancelRecognitionAndInformDelegate(session_id);
-}
-
-void SpeechRecognitionManagerImpl::CancelRecognition(int session_id) {
-  DCHECK(HasPendingRequest(session_id));
-  if (requests_[session_id].is_active)
-    requests_[session_id].recognizer->AbortRecognition();
-  requests_.erase(session_id);
-  if (recording_session_id_ == session_id)
-    recording_session_id_ = 0;
-  if (delegate_.get())
-    delegate_->DoClose(session_id);
-}
-
-void SpeechRecognitionManagerImpl::CancelAllRequestsWithDelegate(
-    InputTagSpeechDispatcherHost* delegate) {
-  SpeechRecognizerMap::iterator it = requests_.begin();
-  while (it != requests_.end()) {
-    if (it->second.delegate == delegate) {
-      CancelRecognition(it->first);
-      // This map will have very few elements so it is simpler to restart.
-      it = requests_.begin();
-    } else {
-      ++it;
-    }
-  }
-}
-
-void SpeechRecognitionManagerImpl::StopRecording(int session_id) {
-  // No pending requests on extension popups.
-  if (!HasPendingRequest(session_id))
-    return;
-
-  requests_[session_id].recognizer->StopAudioCapture();
-}
-
-// -------- SpeechRecognitionEventListener interface implementation. ---------
-
-void SpeechRecognitionManagerImpl::OnRecognitionResult(
-    int session_id, const content::SpeechRecognitionResult& result) {
-  DCHECK(HasPendingRequest(session_id));
-  GetDelegate(session_id)->SetRecognitionResult(session_id, result);
-}
-
-void SpeechRecognitionManagerImpl::OnAudioEnd(int session_id) {
-  if (recording_session_id_ != session_id)
-    return;
-  DCHECK_EQ(recording_session_id_, session_id);
-  DCHECK(HasPendingRequest(session_id));
-  if (!requests_[session_id].is_active)
-    return;
-  recording_session_id_ = 0;
-  GetDelegate(session_id)->DidCompleteRecording(session_id);
-  if (delegate_.get())
-    delegate_->ShowRecognizing(session_id);
-}
-
-void SpeechRecognitionManagerImpl::OnRecognitionEnd(int session_id) {
-  if (!HasPendingRequest(session_id) || !requests_[session_id].is_active)
-    return;
-  GetDelegate(session_id)->DidCompleteRecognition(session_id);
-  requests_.erase(session_id);
-  if (delegate_.get())
-    delegate_->DoClose(session_id);
-}
-
-void SpeechRecognitionManagerImpl::OnSoundStart(int session_id) {
-}
-
-void SpeechRecognitionManagerImpl::OnSoundEnd(int session_id) {
-}
-
-void SpeechRecognitionManagerImpl::OnRecognitionError(
-    int session_id, const content::SpeechRecognitionError& error) {
-  DCHECK(HasPendingRequest(session_id));
-  if (session_id == recording_session_id_)
-    recording_session_id_ = 0;
-  requests_[session_id].is_active = false;
-  if (delegate_.get()) {
-    if (error.code == content::SPEECH_RECOGNITION_ERROR_AUDIO &&
-        error.details == content::SPEECH_AUDIO_ERROR_DETAILS_NO_MIC) {
-      delegate_->ShowMicError(session_id,
-          SpeechRecognitionManagerDelegate::MIC_ERROR_NO_DEVICE_AVAILABLE);
-    } else if (error.code == content::SPEECH_RECOGNITION_ERROR_AUDIO &&
-               error.details == content::SPEECH_AUDIO_ERROR_DETAILS_IN_USE) {
-      delegate_->ShowMicError(session_id,
-          SpeechRecognitionManagerDelegate::MIC_ERROR_DEVICE_IN_USE);
-    } else {
-      delegate_->ShowRecognizerError(session_id, error.code);
-    }
-  }
-}
-
-void SpeechRecognitionManagerImpl::OnAudioStart(int session_id) {
-  DCHECK(HasPendingRequest(session_id));
-  DCHECK_EQ(recording_session_id_, session_id);
-  if (delegate_.get())
-    delegate_->ShowRecording(session_id);
-}
-
-void SpeechRecognitionManagerImpl::OnRecognitionStart(int session_id) {
-}
-
-void SpeechRecognitionManagerImpl::OnEnvironmentEstimationComplete(
-    int session_id) {
-  DCHECK(HasPendingRequest(session_id));
-  DCHECK_EQ(recording_session_id_, session_id);
-}
-
-void SpeechRecognitionManagerImpl::OnAudioLevelsChange(
-    int session_id, float volume, float noise_volume) {
-  DCHECK(HasPendingRequest(session_id));
-  DCHECK_EQ(recording_session_id_, session_id);
-  if (delegate_.get())
-    delegate_->ShowInputVolume(session_id, volume, noise_volume);
-}
-
-void SpeechRecognitionManagerImpl::CancelRecognitionAndInformDelegate(
-    int session_id) {
-  InputTagSpeechDispatcherHost* cur_delegate = GetDelegate(session_id);
-  CancelRecognition(session_id);
-  cur_delegate->DidCompleteRecording(session_id);
-  cur_delegate->DidCompleteRecognition(session_id);
-}
-
-SpeechRecognitionManagerImpl::Request::Request()
-    : is_active(false) {
-}
-
-SpeechRecognitionManagerImpl::Request::~Request() {
+SpeechRecognitionManagerImpl::FSMEventArgs::FSMEventArgs(FSMEvent event_value)
+  : event(event_value),
+    speech_error(content::SPEECH_RECOGNITION_ERROR_NONE) {
+}
+
+SpeechRecognitionManagerImpl::FSMEventArgs::~FSMEventArgs() {
+}
+
+SpeechRecognitionManagerImpl::Session::Session()
+  : id(kSessionIDInvalid),
+    event_listener(NULL),
+    state(STATE_IDLE) {
+}
+
+SpeechRecognitionManagerImpl::Session::~Session() {
 }
 
 }  // namespace speech