【C】语音活性检测

#include "sami_core.h"

// help function
std::vector<uint8_t> loadModelAsBinary(const std::string& path) {
    std::ifstream file(path, std::ios::binary | std::ios::ate);
    std::streamsize size = file.tellg();
    file.seekg(0, std::ios::beg);

    std::vector<uint8_t> buffer(size);
    if(file.read((char*)buffer.data(), size)) { return buffer; }

    return {};
}

SAMICoreFeatureArray* findWantedFeature(SAMICoreFeatureSet* f_set, SAMICorePropertyId feature_id) {
    for(int i = 0; i < f_set->numFeatureTypes; ++i) {
        if(f_set->set[i].featureID == feature_id) { return &(f_set->set[i]); }
    }
    return NULL;
}

// step 1, create handle
SAMICoreHandle handle;
SAMICoreExtractorCreateParam param;
param.sampleRate = sample_rate;
param.numChannel = num_channels;

int ret = SAMICoreCreateHandleByIdentify(&handle, SAMICoreIdentify_Extractor_NNVAD, ¶m);
assert(ret == SAMI_OK);

// step 2, loading model
const std::string model_path = "/path/tob/vad_v1.1.model";
std::vector<uint_8> model_data = loadModelAsBinary(model_path);
assert(model_buf.size() > 0);

SAMICoreProperty coreProperty;
coreProperty.id = SAMICorePropertyId_LoadModelBinary;
 coreProperty.data = model_data.data();
coreProperty.dataLen = model_data.size();
coreProperty.type = SAMICoreDataType_ModelBin;
ret = SAMICoreSetProperty(handle, SAMICorePropertyId_LoadModelBinary, &coreProperty);
assert(ret == SAMI_OK);

// step 3, create input  audio block
SAMICoreAudioBuffer in_audio_buffer;
in_audio_buffer.numberChannels = num_channels;
in_audio_buffer.numberSamples = 0;
in_audio_buffer.isInterleave = 0;
in_audio_buffer.data = new float*[num_channels];

SAMICoreBlock in_block;
in_block.dataType = SAMICoreDataType_AudioBuffer;
in_block.numberAudioData = 1;
in_block.audioData = &in_audio_buffer;

// step 4, process bloock by block
for(;hasAudioSamples();)
{
    updateInputBuffer(in_audio_buffer);
    ret = SAMICoreProcess(handle, &in_block, NULL);
    assert(ret == SAMI_OK);

// step 4.1, get results
    SAMICoreProperty frame_features;
    SAMICoreGetPropertyById(handle, SAMICorePropertyID_FrameFeatures, &frame_features);
    SAMICoreFeatureSet* feature_set = 
                                (SAMICoreFeatureSet*)(frame_features.data);
    if(feature_set != NULL) 
    {
        SAMICoreFeatureArray* result =
            findWantedFeature(feature_set, SAMICorePropertyID_FrameFeature_NNVAD);

        if(result != NULL) {
            float timestamp = result->array[0].time;
            float value = result->array[0].values[0];

            cout << "time at " << timestamp << ", value " << value << endl;
        }
        SAMICoreDestroyProperty(&frame_features);
    }
}

// step 5, remember release resource
ret = SAMICoreDestroyHandle(handle);
assert(ret == SAMI_OK);

delete[] in_audio_buffer.data;

一、创建算法句柄

传入采样率和声道数，通过 SAMICoreCreateHandleByIdentify 创建 handle。

SAMICoreHandle handle;
SAMICoreExtractorCreateParam param;
param.sampleRate = sample_rate;
param.numChannel = num_channels;

int ret = SAMICoreCreateHandleByIdentify(&handle, SAMICoreIdentify_Extractor_NNVAD, ¶m);

二、导入模型

首先读取模型文件到内存，例子中loadModelAsBinary仅供参考；接着，设置模型的信息，并通过SAMICoreSetProperty导入模型

SAMICoreProperty coreProperty;
coreProperty.id = SAMICorePropertyId_LoadModelBinary;
coreProperty.data = model_data.data();
coreProperty.dataLen = model_data.size();
coreProperty.type = SAMICoreDataType_ModelBin;
ret = SAMICoreSetProperty(handle, SAMICorePropertyId_LoadModelBinary, &coreProperty);

三、创建 SAMICoreBlock 用于存放输入和输出

SAMICoreAudioBuffer，用于存放音频数据；SAMICoreBlock，用于存放需要处理的数据。需要注意的是，vad 只检测单声道数据。

SAMICoreAudioBuffer in_buffer;
in_buffer.isInterleave = 0;
in_buffer.numberSamples = 0;
in_buffer.numberChannels = 1;
in_buffer.data = new float*[1];

SAMICoreBlock in_block;
in_block.dataType = SAMICoreDataType_AudioBuffer;
in_block.numberAudioData = 1;
in_block.audioData = &in_buffer;

四、处理音频，并获取结果

更新音频数据的指针，指向正确的内存即可，这样可以避免内存数据的拷贝。

for(;hasAudioSamples();)
{
    in_buffer.data[0] = new_samples_pointer;
    in_buffer.numberSamples = new_samples_length;

    ret = SAMICoreProcess(handle, &in_block, NULL);
    assert(ret == SAMI_OK);

    ....
}

输入是一块一块的音频数据，比如一个 512 大小的音频数据。在调用SAMICoreProcess 后，可以获取当前帧的结果：

for(;hasAudioSamples();)
{
    // .. process

    // step 4.1, get results
    SAMICoreProperty frame_features;
    SAMICoreGetPropertyById(handle, SAMICorePropertyID_FrameFeatures, &frame_features);
    SAMICoreFeatureSet* feature_set = 
                                (SAMICoreFeatureSet*)(frame_features.data);
    if(feature_set != NULL) 
    {
        SAMICoreFeatureArray* result =
            findWantedFeature(feature_set, SAMICorePropertyID_FrameFeature_NNVAD);

        if(result != NULL) {
            float timestamp = result->array[0].time;
            float value = result->array[0].values[0];

            cout << "time at " << timestamp << ", value" << value << endl;
        }
        SAMICoreDestroyProperty(&frame_features);
    }
}

其中：

• 7行， SAMICoreGetPropertyById 通过SAMICorePropertyID_FrameFeatures获取算法所有输出（注：算法可能输出多种类型的特征）

• 12 ~ 13 行，指定 SAMICorePropertyID_FrameFeature_NNVAD，通过 findWantedFeature 辅助函数，获取结果。

• 15~20行，获取时间戳信息，和检测结果

• 21 行，记得调用 SAMICoreDestroyProperty 来释放资源

五、释放资源

释放 handle

ret = SAMICoreDestroyHandle(handle);

此外，还要注意音频数据数据的内存释放（如果有）。例如：

delete[] in_audio_buffer.data;