msb-public
/
PaddleSpeech
mirror of https://github.com/PaddlePaddle/PaddleSpeech
1
0
Code Issues Projects Releases Wiki Activity

fix TTSArmLinux

Browse Source
pull/3030/head
TianYuan 3 years ago
parent 6322cea1a4
commit 96038fb01b
2 changed files with 139 additions and 87 deletions
Show Stats Download Patch File Download Diff File

86
demos/TTSArmLinux/src/Predictor.hpp
Unescape View File

@ -1,7 +1,20 @@
// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <algorithm>
#include <chrono>
#include <iostream>
#include <fstream>
#include <iostream>
#include <memory>
#include <string>
#include <vector>
@ -12,22 +25,26 @@ using namespace paddle::lite_api;
class PredictorInterface {
public:
virtual ~PredictorInterface() = 0;
virtual bool Init(
const std::string &AcousticModelPath,
virtual bool Init(const std::string &AcousticModelPath,
const std::string &VocoderPath,
PowerMode cpuPowerMode,
int cpuThreadNum,
// WAV采样率（必须与模型输出匹配）
// 如果播放速度和音调异常，请修改采样率
// 常见采样率：16000, 24000, 32000, 44100, 48000, 96000
uint32_t wavSampleRate
) = 0;
virtual std::shared_ptr<PaddlePredictor> LoadModel(const std::string &modelPath, int cpuThreadNum, PowerMode cpuPowerMode) = 0;
uint32_t wavSampleRate) = 0;
virtual std::shared_ptr<PaddlePredictor> LoadModel(
const std::string &modelPath,
int cpuThreadNum,
PowerMode cpuPowerMode) = 0;
virtual void ReleaseModel() = 0;
virtual bool RunModel(const std::vector<int64_t> &phones) = 0;
virtual std::unique_ptr<const Tensor> GetAcousticModelOutput(const std::vector<int64_t> &phones) = 0;
virtual std::unique_ptr<const Tensor> GetVocoderOutput(std::unique_ptr<const Tensor> &&amOutput) = 0;
virtual void VocoderOutputToWav(std::unique_ptr<const Tensor> &&vocOutput) = 0;
virtual std::unique_ptr<const Tensor> GetAcousticModelOutput(
const std::vector<int64_t> &phones) = 0;
virtual std::unique_ptr<const Tensor> GetVocoderOutput(
std::unique_ptr<const Tensor> &&amOutput) = 0;
virtual void VocoderOutputToWav(
std::unique_ptr<const Tensor> &&vocOutput) = 0;
virtual void SaveFloatWav(float *floatWav, int64_t size) = 0;
virtual bool IsLoaded() = 0;
virtual float GetInferenceTime() = 0;
@ -48,20 +65,19 @@ PredictorInterface::~PredictorInterface() {}
template <typename WavDataType>
class Predictor : public PredictorInterface {
public:
virtual bool Init(
const std::string &AcousticModelPath,
virtual bool Init(const std::string &AcousticModelPath,
const std::string &VocoderPath,
PowerMode cpuPowerMode,
int cpuThreadNum,
// WAV采样率（必须与模型输出匹配）
// 如果播放速度和音调异常，请修改采样率
// 常见采样率：16000, 24000, 32000, 44100, 48000, 96000
uint32_t wavSampleRate
) override {
uint32_t wavSampleRate) override {
// Release model if exists
ReleaseModel();
acoustic_model_predictor_ = LoadModel(AcousticModelPath, cpuThreadNum, cpuPowerMode);
acoustic_model_predictor_ =
LoadModel(AcousticModelPath, cpuThreadNum, cpuPowerMode);
if (acoustic_model_predictor_ == nullptr) {
return false;
}
@ -80,7 +96,10 @@ public:
ReleaseWav();
}
virtual std::shared_ptr<PaddlePredictor> LoadModel(const std::string &modelPath, int cpuThreadNum, PowerMode cpuPowerMode) override {
virtual std::shared_ptr<PaddlePredictor> LoadModel(
const std::string &modelPath,
int cpuThreadNum,
PowerMode cpuPowerMode) override {
if (modelPath.empty()) {
return nullptr;
}
@ -120,7 +139,8 @@ public:
return true;
}
virtual std::unique_ptr<const Tensor> GetAcousticModelOutput(const std::vector<int64_t> &phones) override {
virtual std::unique_ptr<const Tensor> GetAcousticModelOutput(
const std::vector<int64_t> &phones) override {
auto phones_handle = acoustic_model_predictor_->GetInput(0);
phones_handle->Resize({static_cast<int64_t>(phones.size())});
phones_handle->CopyFromCpu(phones.data());
@ -139,7 +159,8 @@ public:
return am_output_handle;
}
virtual std::unique_ptr<const Tensor> GetVocoderOutput(std::unique_ptr<const Tensor> &&amOutput) override {
virtual std::unique_ptr<const Tensor> GetVocoderOutput(
std::unique_ptr<const Tensor> &&amOutput) override {
auto mel_handle = vocoder_predictor_->GetInput(0);
// [?, 80]
auto dims = amOutput->shape();
@ -161,7 +182,8 @@ public:
return voc_output_handle;
}
virtual void VocoderOutputToWav(std::unique_ptr<const Tensor> &&vocOutput) override {
virtual void VocoderOutputToWav(
std::unique_ptr<const Tensor> &&vocOutput) override {
// 获取输出Tensor的数据
int64_t output_size = 1;
for (auto dim : vocOutput->shape()) {
@ -175,16 +197,13 @@ public:
virtual void SaveFloatWav(float *floatWav, int64_t size) override;
virtual bool IsLoaded() override {
return acoustic_model_predictor_ != nullptr && vocoder_predictor_ != nullptr;
return acoustic_model_predictor_ != nullptr &&
vocoder_predictor_ != nullptr;
}
virtual float GetInferenceTime() override {
return inference_time_;
}
virtual float GetInferenceTime() override { return inference_time_; }
const std::vector<WavDataType> & GetWav() {
return wav_;
}
const std::vector<WavDataType> &GetWav() { return wav_; }
virtual int GetWavSize() override {
return wav_.size() * sizeof(WavDataType);
@ -192,7 +211,8 @@ public:
// 获取WAV持续时间（单位：毫秒）
virtual float GetWavDuration() override {
return static_cast<float>(GetWavSize()) / sizeof(WavDataType) / static_cast<float>(wav_sample_rate_) * 1000;
return static_cast<float>(GetWavSize()) / sizeof(WavDataType) /
static_cast<float>(wav_sample_rate_) * 1000;
}
// 获取RTF（合成时间 / 音频时长）
@ -200,9 +220,7 @@ public:
return GetInferenceTime() / GetWavDuration();
}
virtual void ReleaseWav() override {
wav_.clear();
}
virtual void ReleaseWav() override { wav_.clear(); }
virtual bool WriteWavToFile(const std::string &wavPath) override {
std::ofstream fout(wavPath, std::ios::binary);
@ -216,12 +234,14 @@ public:
header.data_size = GetWavSize();
header.size = sizeof(header) - 8 + header.data_size;
header.sample_rate = wav_sample_rate_;
header.byte_rate = header.sample_rate * header.num_channels * header.bits_per_sample / 8;
header.byte_rate = header.sample_rate * header.num_channels *
header.bits_per_sample / 8;
header.block_align = header.num_channels * header.bits_per_sample / 8;
fout.write(reinterpret_cast<const char *>(&header), sizeof(header));
// 写入wav数据
fout.write(reinterpret_cast<const char*>(wav_.data()), header.data_size);
fout.write(reinterpret_cast<const char *>(wav_.data()),
header.data_size);
fout.close();
return true;
@ -258,9 +278,7 @@ protected:
// 返回值通过模板特化由 WavDataType 决定
inline uint16_t GetWavAudioFormat();
inline float Abs(float number) {
return (number < 0) ? -number : number;
}
inline float Abs(float number) { return (number < 0) ? -number : number; }
protected:
float inference_time_ = 0;

80
demos/TTSArmLinux/src/main.cc
Unescape View File

@ -1,23 +1,48 @@
// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <front/front_interface.h>
#include <gflags/gflags.h>
#include <glog/logging.h>
#include <paddle_api.h>
#include <cstdlib>
#include <iostream>
#include <map>
#include <memory>
#include <string>
#include <map>
#include <glog/logging.h>
#include <gflags/gflags.h>
#include <paddle_api.h>
#include <front/front_interface.h>
#include "Predictor.hpp"
using namespace paddle::lite_api;
DEFINE_string(sentence, "你好，欢迎使用语音合成服务", "Text to be synthesized (Chinese only. English will crash the program.)");
DEFINE_string(
sentence,
"你好，欢迎使用语音合成服务",
"Text to be synthesized (Chinese only. English will crash the program.)");
DEFINE_string(front_conf, "./front.conf", "Front configuration file");
DEFINE_string(acoustic_model, "./models/cpu/fastspeech2_csmsc_arm.nb", "Acoustic model .nb file");
DEFINE_string(vocoder, "./models/cpu/fastspeech2_csmsc_arm.nb", "vocoder .nb file");
DEFINE_string(acoustic_model,
"./models/cpu/fastspeech2_csmsc_arm.nb",
"Acoustic model .nb file");
DEFINE_string(vocoder,
"./models/cpu/fastspeech2_csmsc_arm.nb",
"vocoder .nb file");
DEFINE_string(output_wav, "./output/tts.wav", "Output WAV file");
DEFINE_string(wav_bit_depth, "16", "WAV bit depth, 16 (16-bit PCM) or 32 (32-bit IEEE float)");
DEFINE_string(wav_sample_rate, "24000", "WAV sample rate, should match the output of the vocoder");
DEFINE_string(wav_bit_depth,
"16",
"WAV bit depth, 16 (16-bit PCM) or 32 (32-bit IEEE float)");
DEFINE_string(wav_sample_rate,
"24000",
"WAV sample rate, should match the output of the vocoder");
DEFINE_string(cpu_thread, "1", "CPU thread numbers");
int main(int argc, char *argv[]) {
@ -53,7 +78,7 @@ int main(int argc, char *argv[]) {
// 繁体转简体
std::wstring sentence_simp;
front_inst->Trand2Simp(ws_sentence, sentence_simp);
front_inst->Trand2Simp(ws_sentence, &sentence_simp);
ws_sentence = sentence_simp;
std::string s_sentence;
@ -63,26 +88,28 @@ int main(int argc, char *argv[]) {
// 根据标点进行分句
LOG(INFO) << "Start to segment sentences by punctuation";
front_inst->SplitByPunc(ws_sentence, sentence_part);
front_inst->SplitByPunc(ws_sentence, &sentence_part);
LOG(INFO) << "Segment sentences through punctuation successfully";
// 分句后获取音素id
LOG(INFO) << "Start to get the phoneme and tone id sequence of each sentence";
LOG(INFO)
<< "Start to get the phoneme and tone id sequence of each sentence";
for (int i = 0; i < sentence_part.size(); i++) {
LOG(INFO) << "Raw sentence is: " << ppspeech::wstring2utf8string(sentence_part[i]);
front_inst->SentenceNormalize(sentence_part[i]);
LOG(INFO) << "Raw sentence is: "
<< ppspeech::wstring2utf8string(sentence_part[i]);
front_inst->SentenceNormalize(&sentence_part[i]);
s_sentence = ppspeech::wstring2utf8string(sentence_part[i]);
LOG(INFO) << "After normalization sentence is: " << s_sentence;
if (0 != front_inst->GetSentenceIds(s_sentence, phoneids, toneids)) {
if (0 != front_inst->GetSentenceIds(s_sentence, &phoneids, &toneids)) {
LOG(ERROR) << "TTS inst get sentence phoneids and toneids failed";
return -1;
}
}
LOG(INFO) << "The phoneids of the sentence is: " << limonp::Join(phoneids.begin(), phoneids.end(), " ");
LOG(INFO) << "The toneids of the sentence is: " << limonp::Join(toneids.begin(), toneids.end(), " ");
LOG(INFO) << "The phoneids of the sentence is: "
<< limonp::Join(phoneids.begin(), phoneids.end(), " ");
LOG(INFO) << "The toneids of the sentence is: "
<< limonp::Join(toneids.begin(), toneids.end(), " ");
LOG(INFO) << "Get the phoneme id sequence of each sentence successfully";
@ -99,13 +126,19 @@ int main(int argc, char *argv[]) {
// CPU电源模式
const PowerMode cpuPowerMode = PowerMode::LITE_POWER_HIGH;
if (!predictor->Init(FLAGS_acoustic_model, FLAGS_vocoder, cpuPowerMode, cpuThreadNum, wavSampleRate)) {
if (!predictor->Init(FLAGS_acoustic_model,
FLAGS_vocoder,
cpuPowerMode,
cpuThreadNum,
wavSampleRate)) {
LOG(ERROR) << "predictor init failed" << std::endl;
return -1;
}
std::vector<int64_t> phones(phoneids.size());
std::transform(phoneids.begin(), phoneids.end(), phones.begin(), [](int x) { return static_cast<int64_t>(x); });
std::transform(phoneids.begin(), phoneids.end(), phones.begin(), [](int x) {
return static_cast<int64_t>(x);
});
if (!predictor->RunModel(phones)) {
LOG(ERROR) << "predictor run model failed" << std::endl;
@ -113,7 +146,8 @@ int main(int argc, char *argv[]) {
}
LOG(INFO) << "Inference time: " << predictor->GetInferenceTime() << " ms, "
<< "WAV size (without header): " << predictor->GetWavSize() << " bytes, "
<< "WAV size (without header): " << predictor->GetWavSize()
<< " bytes, "
<< "WAV duration: " << predictor->GetWavDuration() << " ms, "
<< "RTF: " << predictor->GetRTF() << std::endl;

Write Preview
Loading…
Cancel
Save