基于Windows Sdk 与visual C++2008 在微软平台上构架自己的语音识别引擎（适用于windows 2-优快云博客

本文档介绍如何利用Windows SDK和Visual C++2008在微软平台上构建自己的语音识别引擎。项目开源，源码可提供。详细介绍了引擎的框架结构，包括核心类如CSrEngineAlternates、CSampleSRExtension等，并提供了相关的核心源码片段。适用平台包括Windows 2000/xp2003/vista以及CE/Mobile。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

基于Windows Sdk 与visual C++2008 在微软平台上构架自己的语音识别引擎（适用于windows 2000/xp2003/vista windows CE /mobile），本项目开源，源码请留下你们的Email，我给大家发

本人闲来无事，自行开发了一个小型的语音识别引擎，搭建起在微软平台上的语音识别框架服务体系，

鉴于本人个人力量有限，为了将语音识别引擎做的功能更加强悍，强大，

现在将该系统开源，需要源码的请在本人优快云博客下留下EMail，

本系统属于系统框架，搭建起一个语音识别的引擎服务框架，

在微软平台上畅通无阻，

现在将本系统构架公布一下，

并贴出相关核心源码，源码体积为37M，编译后为3M，

适用于windows 2000/xp2003/vista windows CE /mobile

框架头文件简介：

srengalt.h文件此文件包含的语音CSrEngineAlternates类。这实现了接口ISpSRAlternates ，当一个应用程序GetAlternates或识别的结果，将寻找AlternatesCLSID的识别引擎对象，并创建此对象。并返回结果#pragma once#include "stdafx.h"#include "SampleSrEngine.h"#include "resource.h"class ATL_NO_VTABLE CSrEngineAlternates : public CComObjectRootEx<CComMultiThreadModel>, public CComCoClass<CSrEngineAlternates, &CLSID_SampleSREngineAlternates>, public ISpSRAlternates{public: DECLARE_REGISTRY_RESOURCEID(IDR_SRENGALT) DECLARE_PROTECT_FINAL_CONSTRUCT() BEGIN_COM_MAP(CSrEngineAlternates) COM_INTERFACE_ENTRY(ISpSRAlternates) END_COM_MAP()public: STDMETHODIMP GetAlternates( SPPHRASEALTREQUEST *pAltRequest, SPPHRASEALT **ppAlts, ULONG *pcAlts); STDMETHODIMP Commit( SPPHRASEALTREQUEST *pAltRequest, SPPHRASEALT *pAlt, void **ppvResultExtra, ULONG *pcbResultExtra);}; srengext.h 文件此文件包含CSampleSRExtension类。这实现了自定义接口ISampleSRExtension 当一个应用程序开始识别， SAPI的将寻找ExtensionCLSID领域中的引擎对象的指针，并创建该对象，然后创建语音识别的要求。#pragma once#include "stdafx.h" #include "SampleSrEngine.h"#include "resource.h"class ATL_NO_VTABLE CSampleSRExtension : public CComObjectRootEx<CComMultiThreadModel>,public CComCoClass<CSampleSRExtension, &CLSID_SampleSRExtension>,public ISampleSRExtension,public ISpDisplayAlternates,public ISpEnginePronunciation{public:DECLARE_REGISTRY_RESOURCEID(IDR_SRENGEXT)DECLARE_GET_CONTROLLING_UNKNOWN()DECLARE_PROTECT_FINAL_CONSTRUCT() BEGIN_COM_MAP(CSampleSRExtension) COM_INTERFACE_ENTRY(ISampleSRExtension) COM_INTERFACE_ENTRY(ISpDisplayAlternates) COM_INTERFACE_ENTRY(ISpEnginePronunciation)END_COM_MAP() HRESULT FinalConstruct() { / /失败CRecoExt作为一个非累计对象。 / /创建CRecoExt的SAPI if(GetControllingUnknown() == dynamic_cast<ISampleSRExtension *>(this) ) { return E_FAIL; } / /这个接口的处理的SAPI 5.1 。 / /必须QI'd中的FinalConstruct对象并没有释放。 return OuterQueryInterface(IID__ISpPrivateEngineCall, (void **)&m_pEngineCall); } void FinalRelease() { // 不释放IID__ISpPrivateEngineCall这里 } STDMETHODIMP ExamplePrivateEngineCall(void); // 测试方法 // ISpDisplayAlternates 方法 STDMETHODIMP GetDisplayAlternates( const SPDISPLAYPHRASE *pPhrase, ULONG cRequestCount, SPDISPLAYPHRASE **ppCoMemPhrases, ULONG *pcPhrasesReturned); STDMETHODIMP SetFullStopTrailSpace(ULONG ulTrailSpace); // ISpEnginePronunciation 方法 STDMETHODIMP Normalize( LPCWSTR pszWord, LPCWSTR pszLeftContext, LPCWSTR pszRightContext, WORD LangID, SPNORMALIZATIONLIST *pNormalizationList); STDMETHODIMP GetPronunciations( LPCWSTR pszWord, LPCWSTR pszLeftContext, LPCWSTR pszRightContext, WORD LangID, SPWORDPRONUNCIATIONLIST *pEnginePronunciationList);private: _ISpPrivateEngineCall *m_pEngineCall;}; /******************************************************************************* srengobj.h *此文件包含的宣言CSrEngine类。 *本实施ISpSREngine ， ISpSREngine2和ISpObjectWithToken 。 *这是主要识别引擎对象******************************************************************************/#pragma once#include "stdafx.h"#include "SampleSrEngine.h"#include "resource.h"//语音识别对象。每个条目的列表中的一个实例这个类。class CContext{public: CContext * m_pNext; BOOL operator==(SPRECOCONTEXTHANDLE hContext) { return (m_hSapiContext == hContext); } CContext(SPRECOCONTEXTHANDLE hSapiContext) : m_hSapiContext(hSapiContext) {} SPRECOCONTEXTHANDLE m_hSapiContext; };//reco语法存储。每个条目的列表中的一个实例这个类。class CDrvGrammar{public: CDrvGrammar * m_pNext; SPGRAMMARHANDLE m_hSapiGrammar; // 根据 SAPI 创建语法 BOOL m_SLMLoaded; // 语法是否与听写相关 BOOL m_SLMActive; // 词典是否被激活 WCHAR* m_pWordSequenceText; // 词典词表放在缓冲区 ULONG m_cchText; // 字序缓冲区大小 SPTEXTSELECTIONINFO* m_pInfo; // 文字选择字序缓冲区 CDrvGrammar(SPGRAMMARHANDLE hSapiGrammar) : m_hSapiGrammar(hSapiGrammar), m_SLMLoaded(FALSE), m_SLMActive(FALSE), m_pWordSequenceText(NULL), m_cchText(0), m_pInfo(NULL) { } ~CDrvGrammar() { / /释放资源 / /对于每个语法对象将被释放SetWordSequenceData(NULL, 0, NULL). / / SetWordSequenceData和SetTextSelection将释放的内存 / /在这里没有必要释放内存的m_pWordSequenceText和m_pInfo . }#ifdef _WIN32_WCE CDrvGrammar() { } static LONG Compare(const CDrvGrammar *, const CDrvGrammar *) { return 0; }#endif};/ /读取的RecognizeStream线程中的音频数据块。每一组/ /决定如果数据讲话或沉默和价值补充说，此队列。 / /解码器读取这些线程和进程。 / /关键部分是用来使队列线程安全的，和一个事件是用来/ /显示如果缓冲区已空或没有。 / /这非常象roughtly模拟，这样做的特征提取/ /一个线程，并通过功能流的解码器。class CFrameQueue{public: BOOL m_aFrames[100]; // 语音识别返回值 ULONG m_cFrames; ULONG m_ulHeadIndex; HANDLE m_hSpaceAvailEvent; CRITICAL_SECTION m_cs; CFrameQueue() { m_cFrames = 0; m_ulHeadIndex = 0; m_hSpaceAvailEvent = NULL; InitializeCriticalSection(&m_cs); } ~CFrameQueue() { DeleteCriticalSection(&m_cs); } void SetSpaceAvailEvent(HANDLE h) { m_hSpaceAvailEvent = h; } void InsertTail(BOOL b) { EnterCriticalSection(&m_cs); ULONG ulTailIndex = (m_ulHeadIndex + m_cFrames) % sp_countof(m_aFrames); m_aFrames[ulTailIndex] = b; m_cFrames++; if (m_cFrames == sp_countof(m_aFrames)) { ResetEvent(m_hSpaceAvailEvent); } LeaveCriticalSection(&m_cs); } BOOL IsFull() { EnterCriticalSection(&m_cs); BOOL b = (m_cFrames == sp_countof(m_aFrames)); LeaveCriticalSection(&m_cs); return b; } BOOL RemoveHead() { EnterCriticalSection(&m_cs); BOOL b = m_aFrames[m_ulHeadIndex]; m_ulHeadIndex = (m_ulHeadIndex + 1) % sp_countof(m_aFrames); m_cFrames--; SetEvent(m_hSpaceAvailEvent); LeaveCriticalSection(&m_cs); return b; } BOOL HasData() { EnterCriticalSection(&m_cs); ULONG cFrames = m_cFrames; LeaveCriticalSection(&m_cs); return cFrames; }};//我们可以使用CSpBasicQueue信息存储规则class CRuleEntry{public: BOOL operator==(SPRULEHANDLE rh) { return (m_hRule == rh); } CRuleEntry * m_pNext; SPRULEHANDLE m_hRule; // SAPI 规则句柄 BOOL m_fTopLevel; // 显示规则是否被激活 BOOL m_fActive; // 显示识别引擎的设置}; // 语音识别类class ATL_NO_VTABLE CSrEngine : public CComObjectRootEx<CComMultiThreadModel>, public CComCoClass<CSrEngine, &CLSID_SampleSREngine>, public ISpSREngine2, public ISpObjectWithToken, public ISpThreadTask{public: CSrEngine() : m_ulNextGrammarIndex(0), m_cActive(0), m_bPhraseStarted(FALSE), m_bSoundStarted(FALSE), m_hQueueHasRoom(NULL), m_hRequestSync(NULL), m_LangID(0) {}DECLARE_REGISTRY_RESOURCEID(IDR_SRENG)DECLARE_PROTECT_FINAL_CONSTRUCT()BEGIN_COM_MAP(CSrEngine) COM_INTERFACE_ENTRY(ISpSREngine) COM_INTERFACE_ENTRY(ISpSREngine2) COM_INTERFACE_ENTRY(ISpObjectWithToken)END_COM_MAP()private: HANDLE m_hRequestSync; CFrameQueue m_FrameQueue; ULONG m_cBlahBlah; CSpBasicQueue<CDrvGrammar> m_GrammarList; CSpBasicQueue<CContext> m_ContextList; ULONG m_ulNextGrammarIndex; ULONG m_cActive; ULONGLONG m_ullStart; ULONGLONG m_ullEnd; BOOL m_bSoundStarted:1; BOOL m_bPhraseStarted:1; CComPtr<ISpSREngineSite> m_cpSite; CComPtr<ISpThreadControl> m_cpDecoderThread; HANDLE m_hQueueHasRoom; CSpBasicQueue<CRuleEntry> m_RuleList; CComPtr<ISpLexicon> m_cpLexicon; CComPtr<ISpObjectToken> m_cpEngineObjectToken; CComPtr<ISpObjectToken> m_cpUserObjectToken; LANGID m_LangID;public: HRESULT RandomlyWalkRule(SPRECORESULTINFO * pResult, ULONG nWords, ULONGLONG ullAudioPos, ULONG ulAudioSize); HRESULT RecurseWalk(SPSTATEHANDLE hState, SPPATHENTRY * pPath, ULONG * pcTrans); HRESULT WalkCFGRule(SPRECORESULTINFO * pResult, ULONG cRulesActive, BOOL fHypothesis, ULONG nWords, ULONGLONG ullAudioPos, ULONG ulAudioSize); HRESULT WalkSLM(SPRECORESULTINFO * pResult, ULONG cSLMActive, ULONG nWords, ULONGLONG ullAudioPos, ULONG ulAudioSize); HRESULT WalkTextBuffer(void* pvGrammarCookie, SPPATHENTRY * pPath, SPTRANSITIONID hId, ULONG * pcTrans); HRESULT AddEvent(SPEVENTENUM eEvent, ULONGLONG ullStreamPos, WPARAM wParam = 0, LPARAM lParam = 0); HRESULT AddEventString(SPEVENTENUM eEvent, ULONGLONG ulLStreamPos, const WCHAR * psz, WPARAM = 0); HRESULT CreatePhraseFromRule( CRuleEntry * pRule, BOOL fHypothesis, ULONGLONG ullAudioPos, ULONG ulAudioSize, ISpPhraseBuilder** ppPhrase ); CRuleEntry* FindRule( ULONG ulRuleIndex ); CRuleEntry* NextRuleAlt( CRuleEntry * pPriRule, CRuleEntry * pLastRule ); void _CheckRecognition(); void _NotifyRecognition(BOOL fHypothesis, ULONG nWords); HRESULT FinalConstruct(); HRESULT FinalRelease(); STDMETHODIMP SetObjectToken(ISpObjectToken * pToken); STDMETHODIMP GetObjectToken(ISpObjectToken ** ppToken); STDMETHODIMP SetRecoProfile(ISpObjectToken * pProfileToken); STDMETHODIMP SetSite(ISpSREngineSite *pSite); STDMETHODIMP GetInputAudioFormat(const GUID * pSrcFormatId, const WAVEFORMATEX * pSrcWFEX, GUID * pDesiredFormatId, WAVEFORMATEX ** ppCoMemDesiredWFEX); STDMETHODIMP OnCreateRecoContext(SPRECOCONTEXTHANDLE hSAPIRecoContext, void ** ppvDrvCtxt); STDMETHODIMP OnDeleteRecoContext(void * pvDrvCtxt); STDMETHODIMP OnCreateGrammar(void * pvEngineRecoContext, SPGRAMMARHANDLE hSAPIGrammar, void ** ppvEngineGrammar); STDMETHODIMP OnDeleteGrammar(void * pvEngineGrammar); STDMETHODIMP WordNotify(SPCFGNOTIFY Action, ULONG cWords, const SPWORDENTRY * pWords); STDMETHODIMP RuleNotify(SPCFGNOTIFY Action, ULONG cRules, const SPRULEENTRY * pRules); STDMETHODIMP LoadProprietaryGrammar(void * pvEngineGrammar, REFGUID rguidParam, const WCHAR * pszStringParam, const void * pvDataParam, ULONG ulDataSize, SPLOADOPTIONS Options) { return E_NOTIMPL; } STDMETHODIMP UnloadProprietaryGrammar(void * pvEngineGrammar) { return E_NOTIMPL; } STDMETHODIMP SetProprietaryRuleState(void * pvEngineGrammar, const WCHAR * pszName, void * pvReserved, SPRULESTATE NewState, ULONG * pcRulesChanged) { return E_NOTIMPL; } STDMETHODIMP SetProprietaryRuleIdState(void * pvEngineGrammar, DWORD dwRuleId, SPRULESTATE NewState) { return E_NOTIMPL; } /由于这个引擎不支持专有的语法，我们并不需要执行 / /此方法不仅仅是返回S_OK 。注意执行不返回 E_NOTIMPL 。 / /仅仅返回S_OK ，并忽略这个数据如果您不需要它执行专有语法。 STDMETHODIMP SetGrammarState(void * pvEngineGrammar, SPGRAMMARSTATE eGrammarState) { return S_OK; } STDMETHODIMP SetContextState(void * pvEngineContxt, SPCONTEXTSTATE eCtxtState) { return S_OK; } // 字典方法 STDMETHODIMP LoadSLM(void * pvEngineGrammar, const WCHAR * pszTopicName); STDMETHODIMP UnloadSLM(void * pvEngineGrammar); STDMETHODIMP SetSLMState(void * pvEngineGrammar, SPRULESTATE NewState); STDMETHODIMP IsPronounceable(void *pDrvGrammar, const WCHAR *pszWord, SPWORDPRONOUNCEABLE * pWordPronounceable); STDMETHODIMP SetWordSequenceData(void * pvEngineGrammar, const WCHAR * pText, ULONG cchText, const SPTEXTSELECTIONINFO * pInfo); STDMETHODIMP SetTextSelection(void * pvEngineGrammar, const SPTEXTSELECTIONINFO * pInfo); STDMETHODIMP SetAdaptationData(void * pvEngineCtxtCookie, const WCHAR * pText, const ULONG cch); STDMETHODIMP SetPropertyNum( SPPROPSRC eSrc, void* pvSrcObj, const WCHAR* pName, LONG lValue ); STDMETHODIMP GetPropertyNum( SPPROPSRC eSrc, void* pvSrcObj, const WCHAR* pName, LONG * plValue ); STDMETHODIMP SetPropertyString( SPPROPSRC eSrc, void* pvSrcObj, const WCHAR* pName, const WCHAR* pValue ); STDMETHODIMP GetPropertyString( SPPROPSRC eSrc, void* pvSrcObj, const WCHAR* pName, __deref_out_opt WCHAR** ppCoMemValue ); // 语音识别方法 STDMETHODIMP RecognizeStream(REFGUID rguidFmtId, const WAVEFORMATEX * pWaveFormatEx, HANDLE hRequestSync, HANDLE hDataAvailable, HANDLE hExit, BOOL fNewAudioStream, BOOL fRealTimeAudio, ISpObjectToken * pAudioObjectToken); STDMETHODIMP PrivateCall(void * pvEngineContext, void * pCallFrame, ULONG ulCallFrameSize); STDMETHODIMP PrivateCallEx(void * pvEngineContext, const void * pInCallFrame, ULONG ulCallFrameSize, void ** ppvCoMemResponse, ULONG * pcbResponse); // 语音识别线程 STDMETHODIMP InitThread( void * pvTaskData, HWND hwnd ) { return S_OK; } LRESULT STDMETHODCALLTYPE WindowMessage( void *pvTaskData, HWND hWnd, UINT Msg, WPARAM wParam, LPARAM lParam ) { return E_UNEXPECTED; } STDMETHODIMP ThreadProc( void *pvTaskData, HANDLE hExitThreadEvent, HANDLE hNotifyEvent, HWND hwndWorker, volatile const BOOL * pfContinueProcessing ); // 语音引擎方法 STDMETHODIMP PrivateCallImmediate( void *pvEngineContext, const void *pInCallFrame, ULONG ulInCallFrameSize, void **ppvCoMemResponse, ULONG *pulResponseSize); STDMETHODIMP SetAdaptationData2( void *pvEngineContext, __in_ecount(cch) const WCHAR *pAdaptationData, const ULONG cch, LPCWSTR pTopicName, SPADAPTATIONSETTINGS eSettings, SPADAPTATIONRELEVANCE eRelevance); STDMETHODIMP SetGrammarPrefix( void *pvEngineGrammar, __in_opt LPCWSTR pszPrefix, BOOL fIsPrefixRequired); STDMETHODIMP SetRulePriority( SPRULEHANDLE hRule, void *pvClientRuleContext, int nRulePriority); STDMETHODIMP EmulateRecognition( ISpPhrase *pPhrase, DWORD dwCompareFlags); STDMETHODIMP SetSLMWeight( void *pvEngineGrammar, float flWeight); STDMETHODIMP SetRuleWeight( SPRULEHANDLE hRule, void *pvClientRuleContext, float flWeight); STDMETHODIMP SetTrainingState( BOOL fDoingTraining, BOOL fAdaptFromTrainingData); STDMETHODIMP ResetAcousticModelAdaptation( void); STDMETHODIMP OnLoadCFG( void *pvEngineGrammar, const SPBINARYGRAMMAR *pGrammarData, ULONG ulGrammarID); STDMETHODIMP OnUnloadCFG( void *pvEngineGramma