|
#include <iostream> |
|
#include <torch/torch.h> |
|
#include <torch/script.h> |
|
#include <string> |
|
#include <vector> |
|
#include <locale> |
|
#include <codecvt> |
|
#include <direct.h> |
|
#include <fstream> |
|
typedef int64_t int64; |
|
namespace Shirakana { |
|
|
|
struct WavHead { |
|
char RIFF[4]; |
|
long int size0; |
|
char WAVE[4]; |
|
char FMT[4]; |
|
long int size1; |
|
short int fmttag; |
|
short int channel; |
|
long int samplespersec; |
|
long int bytepersec; |
|
short int blockalign; |
|
short int bitpersamples; |
|
char DATA[4]; |
|
long int size2; |
|
}; |
|
|
|
int conArr2Wav(int64 size, int16_t* input, const char* filename) { |
|
WavHead head = { {'R','I','F','F'},0,{'W','A','V','E'},{'f','m','t',' '},16, |
|
1,1,22050,22050 * 2,2,16,{'d','a','t','a'}, |
|
0 }; |
|
head.size0 = size * 2 + 36; |
|
head.size2 = size * 2; |
|
std::ofstream ocout; |
|
char* outputData = (char*)input; |
|
ocout.open(filename, std::ios::out | std::ios::binary); |
|
ocout.write((char*)&head, 44); |
|
ocout.write(outputData, (int32_t)(size * 2)); |
|
ocout.close(); |
|
return 0; |
|
} |
|
|
|
inline std::wstring to_wide_string(const std::string& input) |
|
{ |
|
std::wstring_convert<std::codecvt_utf8<wchar_t>> converter; |
|
return converter.from_bytes(input); |
|
} |
|
|
|
inline std::string to_byte_string(const std::wstring& input) |
|
{ |
|
std::wstring_convert<std::codecvt_utf8<wchar_t>> converter; |
|
return converter.to_bytes(input); |
|
} |
|
} |
|
|
|
#define val const auto |
|
int main() |
|
{ |
|
torch::jit::Module Vits; |
|
std::string buffer; |
|
std::vector<int64> text; |
|
std::vector<int16_t> data; |
|
while(true) |
|
{ |
|
while (true) |
|
{ |
|
std::cin >> buffer; |
|
if (buffer == "end") |
|
return 0; |
|
if(buffer == "model") |
|
{ |
|
std::cin >> buffer; |
|
Vits = torch::jit::load(buffer); |
|
continue; |
|
} |
|
if (buffer == "endinfer") |
|
{ |
|
Shirakana::conArr2Wav(data.size(), data.data(), "temp\\tmp.wav"); |
|
data.clear(); |
|
std::cout << "endofinfe"; |
|
continue; |
|
} |
|
if (buffer == "line") |
|
{ |
|
std::cin >> buffer; |
|
while (buffer.find("endline")==std::string::npos) |
|
{ |
|
text.push_back(std::atoi(buffer.c_str())); |
|
std::cin >> buffer; |
|
} |
|
val InputTensor = torch::from_blob(text.data(), { 1,static_cast<int64>(text.size()) }, torch::kInt64); |
|
std::array<int64, 1> TextLength{ static_cast<int64>(text.size()) }; |
|
val InputTensor_length = torch::from_blob(TextLength.data(), { 1 }, torch::kInt64); |
|
std::vector<torch::IValue> inputs; |
|
inputs.push_back(InputTensor); |
|
inputs.push_back(InputTensor_length); |
|
if (buffer.length() > 7) |
|
{ |
|
std::array<int64, 1> speakerIndex{ (int64)atoi(buffer.substr(7).c_str()) }; |
|
inputs.push_back(torch::from_blob(speakerIndex.data(), { 1 }, torch::kLong)); |
|
} |
|
val output = Vits.forward(inputs).toTuple()->elements()[0].toTensor().multiply(32276.0F); |
|
val outputSize = output.sizes().at(2); |
|
val floatOutput = output.data_ptr<float>(); |
|
int16_t* outputTmp = (int16_t*)malloc(sizeof(float) * outputSize); |
|
if (outputTmp == nullptr) { |
|
throw std::exception("内存不足"); |
|
} |
|
for (int i = 0; i < outputSize; i++) { |
|
*(outputTmp + i) = (int16_t) * (floatOutput + i); |
|
} |
|
data.insert(data.end(), outputTmp, outputTmp+outputSize); |
|
free(outputTmp); |
|
text.clear(); |
|
std::cout << "endofline"; |
|
} |
|
} |
|
} |
|
|
|
} |
