基于GMM的说话人识别（C代码）

/** @file ******************************************************************************** <PRE> 模块名: mfcc参数的提取 文件名: mfcc.c 相关文件: mfcc.h 文件实现功能: mfcc提取相关函数的实现 作者: Dake 版本: V2010.09.01 编程方式: ANSI C语言编程 授权方式: Copyright(C) Dake 联系方式: [email protected] 生成日期: 2010-06-22 -------------------------------------------------------------------------------- 多线程安全性: <是/否>[，说明] 异常时安全性: <是/否>[，说明] -------------------------------------------------------------------------------- 备注: <其它说明> -------------------------------------------------------------------------------- 修改记录: 日 期 版本 修改人 修改内容 YYYY/MM/DD X.Y <作者或修改者名> <修改内容> </PRE> *******************************************************************************/ #include "mfcc.h" static void PreEmphasis(BYTE * pRawVoice, int iSampleNum, double * pDiffVoice); // 预增强 static double * JudgeTrainFrame2(double * pDiffVoice, int iSampleNum, int train_frame_num); // 判断是否为静寂语音 static BOOL FrameMultiHamming(double * pDiffVoice, VOICEFRAME * pVoiceFrame, int train_frame_num); static void FrameToMFCC(VOICEFRAME * pVoiceFrame, double ** MFCC, int train_frame_num); static double HalfFrameEnergy(double * pDiffVoice); // 计算半帧能量 /** * @brief 三角带通滤波器数组 */ static double melf16[239] = { //0 [3] 1.4254, 1.2616, 0.0445, //1 [3] 0.7384, 1.9555, 0.9103, //2 [3] 1.0897, 1.8484, 0.8501, //3 [5] 0.1516, 1.1499, 1.9082, 1.0167, 0.1705, //4 [5] 0.0918, 0.9833, 1.8295, 1.3652, 0.5970, //5 [5] 0.6348, 1.4030, 1.8627, 1.1593, 0.4845, //6 [7] 0.1373, 0.8407, 1.5155, 1.8359, 1.2116, 0.6099, 0.0291, //7 [7] 0.1641, 0.7884, 1.3901, 1.9709, 1.4679, 0.9249, 0.3991, //8 [7] 0.5321, 1.0751, 1.6009, 1.8894, 1.3948, 0.9144, 0.4476, //9 [9] 0.1106, 0.6052, 1.0856, 1.5524, 1.9934, 1.5513, 1.1206, 0.7008, 0.2912, //10 [11] 0.0066, 0.4487, 0.8794, 1.2992, 1.7088, 1.8915, 1.5012, 1.1198, 0.7469, 0.3822, 0.0253, //11 [12] 0.1085, 0.4988, 0.8802, 1.2531, 1.6178, 1.9747, 1.6759, 1.3336, 0.9983, 0.6696, 0.3472, 0.0309, //12 [12] 0.3241, 0.6664, 1.0017, 1.3304, 1.6528, 1.9691, 1.7205, 1.4158, 1.1166, 0.8226, 0.5338, 0.2498, //13 [14] 0.2795, 0.5842, 0.8834, 1.1774, 1.4662, 1.7502, 1.9707, 1.6961, 1.4260, 1.1601, 0.8985, 0.6409, 0.3872, 0.1373, //14 [17] 0.0293, 0.3039, 0.5740, 0.8399, 1.1015, 1.3591, 1.6128, 1.8627, 1.8911, 1.6485, 1.4094, 1.1737, 0.9412, 0.7119, 0.4858, 0.2627, 0.0425, //15 [18] 0.1089, 0.3515, 0.5906, 0.8263, 1.0588, 1.2881, 1.5142, 1.7373, 1.9575, 1.8252, 1.6107, 1.3989, 1.1898, 0.9833, 0.7793, 0.5777, 0.3786, 0.1818, //16 [20] 0.1748, 0.3893, 0.6011, 0.8102, 1.0167, 1.2207, 1.4223, 1.6214, 1.8182, 1.9874, 1.7951, 1.6051, 1.4172, 1.2314, 1.0476, 0.8659, 0.6861, 0.5082, 0.3322, 0.1580, //17 [24] 0.0126, 0.2049, 0.3949, 0.5828, 0.7686, 0.9524, 1.1341, 1.3139, 1.4918, 1.6678, 1.8420, 1.9857, 1.8151, 1.6462, 1.4791, 1.3136, 1.1497, 0.9874, 0.8267, 0.6675, 0.5098, 0.3536, 0.1988, 0.0455, //18 [27] 0.0143, 0.1849, 0.3538, 0.5209, 0.6864, 0.8503, 1.0126, 1.1733, 1.3325, 1.4902, 1.6464, 1.8012, 1.9545, 1.8936, 1.7430, 1.5938, 1.4459, 1.2993, 1.1540, 1.0100, 0.8672, 0.7256, 0.5852, 0.4459, 0.3078, 0.1709, 0.0350, //19 [29] 0.1064, 0.2570, 0.4062, 0.5541, 0.7007, 0.8460, 0.9900, 1.1328, 1.2744, 1.4148, 1.5541, 1.6922, 1.8291, 1.9650, 1.9003, 1.7667, 1.6341, 1.5025, 1.3720, 1.2425, 1.1140, 0.9865, 0.8599, 0.7343, 0.6097, 0.4859, 0.3631, 0.2412, 0.1202 }; /** * @brief 三角带通滤波器数组下标索引，表示（滤波器数组中的起始位置，最小列号，最大列号） */ static int indexM16[D][3] = { {0, 1, 3}, {3, 2, 4}, {6, 4, 6}, {9, 5, 9}, {14, 7, 11}, {19, 10, 14}, {24, 12, 18}, {31, 15, 21}, {38, 19, 25}, {45, 22, 30}, {54, 26, 36}, {65, 31, 42}, {77, 37, 48}, {89, 43, 56}, {103, 49, 65}, {120, 57, 74}, {138, 66, 85}, {158, 75, 98}, {182, 86, 112}, {209, 99, 127}, }; /** @function ******************************************************************************** <PRE> 函数名: PreEmphasis() 功能: 预增强，对原始语音数据做差分 用法: 参数: [IN] pRawVoice: 原始语音数据 [IN] iSampleNum: 语音样本数 [OUT] pDiffVoice: 差分后的数据 返回: 调用: 主调函数: voiceToMFCC() </PRE> *******************************************************************************/ static void PreEmphasis(BYTE * pRawVoice, int iSampleNum, double * pDiffVoice) { int i; for (i = 0; i < iSampleNum; ++i) { pDiffVoice[i] = (pRawVoice[i] - 128) * 0.0078125; } } /** @function ******************************************************************************** <PRE> 函数名: JudgeTrainFrame2() 功能: 判断预增强后的数据是否可以作为训练帧（去除静寂语音的多段连接起来） 用法: 参数: [IN] pDiffVoice: 增强（差分）后的语音数据 [IN] iSampleNum: 语音样本数 [IN] train_frame_num: 训练帧帧数 返回: 非空: 可以作为训练帧的信号首地址 NULL: 不可作为训练帧 调用: HalfFrameEnergy() 主调函数: voiceToMFCC() 备注: 返回的内存需要用户在voiceToMFCC()中自行释放 </PRE> *******************************************************************************/ static double * JudgeTrainFrame2(double * pDiffVoice, int iSampleNum, int train_frame_num) { int len = 0; int iNum; // 合格的语音帧数 int iFrame; // 帧序号 int iFrameNum = (iSampleNum>>7) - 1; // 帧总数=2*(样本数/256)-1 double PriorHalfFrame = 0; // 前半帧能量(前128B) double NextHalfFrame = 0; // 后半帧能量(后128B) double FrameEnergy = 0; // 一帧的累积能量 double * pTranDiff = NULL; // 合格语音 double * pSave; // 合格语音存储地址 double * pDiffHead = pDiffVoice; double * pDiffNext = pDiffVoice; if (iFrameNum < train_frame_num) { return NULL; } pTranDiff = (double *)calloc((train_frame_num/2 + 1)*FRAME_LEN, sizeof(double)); if (!pTranDiff) { return NULL; } pSave = pTranDiff; PriorHalfFrame = HalfFrameEnergy(pDiffHead); for (iFrame=1, iNum=0; iFrame<iFrameNum && iNum<train_frame_num; ++iFrame) { pDiffNext += (FRAME_LEN/2); // 指向iFrame号帧的后半帧 NextHalfFrame = HalfFrameEnergy(pDiffNext); FrameEnergy = PriorHalfFrame + NextHalfFrame; if (FrameEnergy < SILENCE_VALUE) // 是静寂语音 { if (iFrameNum-iFrame < train_frame_num-iNum) // 剩下的帧不足以作为训练帧 { free(pTranDiff); return NULL; } len = pDiffNext - pDiffHead - FRAME_LEN/2; if (len > 0) { memcpy(pSave, pDiffHead, len * sizeof(double)); pSave += len; } pDiffNext += FRAME_LEN/2; // 重新定位可用作训练的帧 pDiffHead= pDiffNext; NextHalfFrame = HalfFrameEnergy(pDiffHead); } else { ++iNum; } PriorHalfFrame = NextHalfFrame; } // end: for(;;) len = pDiffNext - pDiffHead + FRAME_LEN/2; if (len > 0) { memcpy(pSave, pDiffHead, len * sizeof(double)); } return pTranDiff; } /** @function