add c++ inference based on libtorch

examples/cpp_libtorch/ReadMe

@@ -0,0 +1,35 @@
+This is the source code for Silero-VAD 5.1 in C++, based on LibTorch.
+The primary implementation is the CPU version, and you should compare its results with the Python version.
+
+In addition, Batch and CUDA inference options are also available if you want to explore further. 
+Note that when using batch inference, the speech probabilities might slightly differ from the standard version, likely due to differences in caching. 
+Unlike processing individual inputs, batch inference may not be able to use the cache from previous chunks. 
+Nevertheless, batch inference provides significantly faster processing. 
+For optimal performance, carefully adjust the threshold when using batch inference.
+
+#Requirements:
+GCC 11.4.0 (GCC >= 5.1)
+LibTorch 1.13.0(Other versions are also acceptable)
+
+#Download Libtorch:
+   #cpu
+   $wget https://download.pytorch.org/libtorch/cpu/libtorch-shared-with-deps-1.13.0%2Bcpu.zip
+   $unzip libtorch-shared-with-deps-1.13.0+cpu.zip 
+
+   #cuda
+   $wget https://download.pytorch.org/libtorch/cu116/libtorch-shared-with-deps-1.13.0%2Bcu116.zip
+   $unzip libtorch-shared-with-deps-1.13.0+cu116.zip
+
+#complie:
+   #cpu
+   $g++ main.cc silero_torch.cc -I ./libtorch/include/ -I ./libtorch/include/torch/csrc/api/include -L ./libtorch/lib/ -ltorch -ltorch_cpu -lc10 -Wl,-rpath,./libtorch/lib/ -o silero -std=c++14 -D_GLIBCXX_USE_CXX11_ABI=0
+   #cuda
+   $g++ main.cc silero_torch.cc -I ./libtorch/include/ -I ./libtorch/include/torch/csrc/api/include -L ./libtorch/lib/ -ltorch -ltorch_cuda -ltorch_cpu -lc10 -Wl,-rpath,./libtorch/lib/ -o silero -std=c++14 -D_GLIBCXX_USE_CXX11_ABI=0 -DUSE_GPU
+
+   #option to add
+   -DUSE_BATCH
+   -DUSE_GPU
+
+# Run:
+./silero aepyx.wav 0.5	#The sample file 'aepyx.wav' is part of the Voxconverse dataset.
+			#aepyx.wav : 16kHz, 16-bit

examples/cpp_libtorch/main.cc

@@ -0,0 +1,51 @@
+#include <iostream>
+#include "silero_torch.h"
+#include "wav.h"
+
+int main(int argc, char* argv[]) {
+
+	if(argc != 3){
+		std::cerr<<"Usage : "<<argv[0]<<" <wav.path> threshold"<<std::endl;
+		std::cerr<<"Usage : "<<argv[0]<<" sample.wav 0.38"<<std::endl;
+		return 1;
+	}
+
+	std::string wav_path = argv[1];
+	float threshold = std::stof(argv[2]);
+
+
+	//Load Model
+	std::string model_path = "../../src/silero_vad/data/silero_vad.jit";
+	silero::VadIterator vad(model_path);
+        vad.threshold=threshold;
+        vad.min_speech_duration_ms=255;
+        vad.max_duration_merge_ms=300;
+        vad.print_as_samples=true;	//if true, it prints time-stamp with sample numbers.
+					//(Default:false)
+
+	// Read wav
+	wav::WavReader wav_reader(wav_path); 
+	std::vector<float> input_wav(wav_reader.num_samples());
+
+	for (int i = 0; i < wav_reader.num_samples(); i++)
+	{
+		input_wav[i] = static_cast<float>(*(wav_reader.data() + i));
+	}
+
+	vad.SpeechProbs(input_wav);
+
+	std::vector<silero::Interval> speeches = vad.GetSpeechTimestamps();
+	for(const auto& speech : speeches){
+		if(vad.print_as_samples){
+			std::cout<<"{'start': "<<static_cast<int>(speech.start)<<", 'end': "<<static_cast<int>(speech.end)<<"}"<<std::endl;
+		}
+		else{
+			std::cout<<"{'start': "<<speech.start<<", 'end': "<<speech.end<<"}"<<std::endl;
+		}
+	}	
+
+
+	return 0;
+	}
+
+

examples/cpp_libtorch/silero_torch.cc

@@ -0,0 +1,283 @@
+//Author      : Nathan Lee
+//Created On  : 2024-11-18
+//Description : silero 5.1 system for torch-script(c++).
+//Version     : 1.0
+//Contact     : junghan4242@gmail.com
+
+
+#include "silero_torch.h"
+
+namespace silero {
+
+	VadIterator::VadIterator(const std::string &model_path, float threshold, int sample_rate, int window_size_ms, int speech_pad_ms, int min_silence_duration_ms, int min_speech_duration_ms, int max_duration_merge_ms, bool print_as_samples)
+		:sample_rate(sample_rate), threshold(threshold), speech_pad_ms(speech_pad_ms), min_silence_duration_ms(min_silence_duration_ms), min_speech_duration_ms(min_speech_duration_ms), max_duration_merge_ms(max_duration_merge_ms), print_as_samples(print_as_samples)
+	{
+		init_torch_model(model_path);
+		init_engine(window_size_ms);
+	}
+	VadIterator::~VadIterator(){
+	}
+
+
+	void VadIterator::SpeechProbs(std::vector<float>& input_wav){
+		// Set the sample rate (must match the model's expected sample rate)
+		// Process the waveform in chunks of 512 samples
+		int num_samples = input_wav.size();
+		int num_chunks = num_samples / window_size_samples;
+		int remainder_samples = num_samples % window_size_samples;
+
+		total_sample_size += num_samples;
+
+		torch::Tensor output;
+		std::vector<torch::Tensor> chunks;
+
+		for (int i = 0; i < num_chunks; i++) {
+
+			float* chunk_start = input_wav.data() + i *window_size_samples;
+			torch::Tensor chunk = torch::from_blob(chunk_start, {1,window_size_samples}, torch::kFloat32);
+			//std::cout<<"chunk size : "<<chunk.sizes()<<std::endl;
+			chunks.push_back(chunk);
+
+
+			if(i==num_chunks-1 && remainder_samples>0){//마지막 chunk && 나머지가 존재
+				int remaining_samples = num_samples - num_chunks * window_size_samples;
+				//std::cout<<"Remainder size : "<<remaining_samples;
+				float* chunk_start_remainder = input_wav.data() + num_chunks *window_size_samples;
+
+				torch::Tensor remainder_chunk = torch::from_blob(chunk_start_remainder, {1,remaining_samples},
+						torch::kFloat32);
+				// Pad the remainder chunk to match window_size_samples
+				torch::Tensor padded_chunk = torch::cat({remainder_chunk, torch::zeros({1, window_size_samples
+							- remaining_samples}, torch::kFloat32)}, 1);
+				//std::cout<<", padded_chunk size : "<<padded_chunk.size(1)<<std::endl;
+
+				chunks.push_back(padded_chunk);
+			}
+		}
+
+		if (!chunks.empty()) {
+
+#ifdef USE_BATCH
+			torch::Tensor batched_chunks = torch::stack(chunks);  // Stack all chunks into a single tensor
+			//batched_chunks = batched_chunks.squeeze(1);
+			batched_chunks = torch::cat({batched_chunks.squeeze(1)});
+
+#ifdef USE_GPU
+			batched_chunks = batched_chunks.to(at::kCUDA);        // Move the entire batch to GPU once
+#endif
+			// Prepare input for model
+			std::vector<torch::jit::IValue> inputs;
+			inputs.push_back(batched_chunks);  // Batch of chunks
+			inputs.push_back(sample_rate);     // Assuming sample_rate is a valid input for the model
+
+			// Run inference on the batch
+			torch::NoGradGuard no_grad;
+			torch::Tensor output = model.forward(inputs).toTensor();
+#ifdef USE_GPU
+			output = output.to(at::kCPU);      // Move the output back to CPU once
+#endif
+			// Collect output probabilities
+			for (int i = 0; i < chunks.size(); i++) {
+				float output_f = output[i].item<float>();
+				outputs_prob.push_back(output_f);
+				//std::cout << "Chunk " << i << " prob: " << output_f<< "\n";
+			}
+#else
+
+			std::vector<torch::Tensor> outputs;
+			torch::Tensor batched_chunks = torch::stack(chunks);
+#ifdef USE_GPU
+			batched_chunks = batched_chunks.to(at::kCUDA);
+#endif
+			for (int i = 0; i < chunks.size(); i++) {
+				torch::NoGradGuard no_grad;
+				std::vector<torch::jit::IValue> inputs;
+				inputs.push_back(batched_chunks[i]);
+				inputs.push_back(sample_rate);
+
+				torch::Tensor output = model.forward(inputs).toTensor();
+				outputs.push_back(output);
+			}
+			torch::Tensor all_outputs = torch::stack(outputs);
+#ifdef USE_GPU
+			all_outputs = all_outputs.to(at::kCPU);
+#endif
+			for (int i = 0; i < chunks.size(); i++) {
+				float output_f = all_outputs[i].item<float>();
+				outputs_prob.push_back(output_f);
+			}
+
+
+
+#endif
+
+		}
+
+
+	}
+
+
+	std::vector<Interval> VadIterator::GetSpeechTimestamps() {
+		std::vector<Interval> speeches = DoVad();
+
+#ifdef USE_BATCH
+		//When you use BATCH inference. You would better use 'mergeSpeeches' function to arrage time stamp.
+		//It could be better get reasonable output because of distorted probs.
+		duration_merge_samples = sample_rate * max_duration_merge_ms / 1000;
+		std::vector<Interval> speeches_merge = mergeSpeeches(speeches, duration_merge_samples);
+		if(!print_as_samples){
+			for (auto& speech : speeches_merge) { //samples to second
+				speech.start /= sample_rate;
+				speech.end /= sample_rate;
+			}
+		}
+
+		return speeches_merge;
+#else
+
+		if(!print_as_samples){
+			for (auto& speech : speeches) { //samples to second
+				speech.start /= sample_rate;
+				speech.end /= sample_rate;
+			}
+		}
+
+		return speeches;
+
+#endif
+
+	}
+
+	void VadIterator::init_engine(int window_size_ms) {
+		min_silence_samples = sample_rate * min_silence_duration_ms / 1000;
+		speech_pad_samples = sample_rate * speech_pad_ms / 1000;
+		window_size_samples = sample_rate / 1000 * window_size_ms;
+		min_speech_samples = sample_rate * min_speech_duration_ms / 1000;
+	}
+
+	void VadIterator::init_torch_model(const std::string& model_path) {
+		at::set_num_threads(1);
+		model = torch::jit::load(model_path);
+
+#ifdef USE_GPU
+		if (!torch::cuda::is_available()) {
+			std::cout<<"CUDA is not available! Please check your GPU settings"<<std::endl;
+			throw std::runtime_error("CUDA is not available!");
+			model.to(at::Device(at::kCPU));    
+
+		} else {
+			std::cout<<"CUDA available! Running on '0'th GPU"<<std::endl;
+			model.to(at::Device(at::kCUDA, 0));        //select 0'th machine 
+		}
+#endif
+
+
+		model.eval();
+		torch::NoGradGuard no_grad;
+		std::cout << "Model loaded successfully"<<std::endl;
+	}
+
+	void VadIterator::reset_states() {
+		triggered = false;
+		current_sample = 0;
+		temp_end = 0;
+		outputs_prob.clear();
+		model.run_method("reset_states");
+		total_sample_size = 0;
+	}
+
+	std::vector<Interval> VadIterator::DoVad() {
+		std::vector<Interval> speeches;
+
+		for (size_t i = 0; i < outputs_prob.size(); ++i) {
+			float speech_prob = outputs_prob[i];
+			//std::cout << speech_prob << std::endl;
+			//std::cout << "Chunk " << i << " Prob: " << speech_prob << "\n";
+			//std::cout << speech_prob << " ";
+			current_sample += window_size_samples;
+
+			if (speech_prob >= threshold && temp_end != 0) {
+				temp_end = 0;
+			}
+
+			if (speech_prob >= threshold && !triggered) {
+				triggered = true;
+				Interval segment;
+				segment.start = std::max(static_cast<int>(0), current_sample - speech_pad_samples - window_size_samples);
+				speeches.push_back(segment);
+				continue;
+			}
+
+			if (speech_prob < threshold - 0.15f && triggered) {
+				if (temp_end == 0) {
+					temp_end = current_sample;
+				}
+
+				if (current_sample - temp_end < min_silence_samples) {
+					continue;
+				} else {
+					Interval& segment = speeches.back();
+					segment.end = temp_end + speech_pad_samples - window_size_samples;
+					temp_end = 0;
+					triggered = false;
+				}
+			}
+		}
+
+		if (triggered) { //만약 낮은 확률을 보이다가  마지막프레임 prbos만 딱 확률이 높게 나오면 위에서 triggerd = true 메핑과 동시에  segment start가 돼서 문제가 될것 같은데? start = end 같은값? 후처리가 있으니 문제가 없으려나?
+			std::cout<<"when last triggered is keep working until last Probs"<<std::endl;
+			Interval& segment = speeches.back();
+			segment.end = total_sample_size;  // 현재 샘플을 마지막 구간의 종료 시간으로 설정
+			triggered = false;  // VAD 상태 초기화
+		}
+
+		speeches.erase(
+                		std::remove_if(
+                        speeches.begin(),
+                        speeches.end(),
+                        [this](const Interval& speech) {
+                        return ((speech.end - this->speech_pad_samples) - (speech.start + this->speech_pad_samples) < min_speech_samples);
+			//min_speech_samples is 4000samples(0.25sec)
+			//여기서 포인트!! 계산 할때는 start,end sample에'speech_pad_samples' 사이즈를 추가한후 길이를 측정함. 
+                        }
+                ),
+                speeches.end()
+              );
+
+
+		//std::cout<<std::endl;
+		//std::cout<<"outputs_prob.size : "<<outputs_prob.size()<<std::endl;
+
+		reset_states();
+		return speeches;
+	}
+
+	std::vector<Interval> VadIterator::mergeSpeeches(const std::vector<Interval>& speeches, int duration_merge_samples) {
+		std::vector<Interval> mergedSpeeches;
+
+		if (speeches.empty()) {
+			return mergedSpeeches; // 빈 벡터 반환
+		}
+
+		// 첫 번째 구간으로 초기화
+		Interval currentSegment = speeches[0];
+
+		for (size_t i = 1; i < speeches.size(); ++i) {	//첫번째 start,end 정보 건너뛰기. 그래서 i=1부터
+			// 두 구간의 차이가 threshold(duration_merge_samples)보다 작은 경우, 합침
+			if (speeches[i].start - currentSegment.end < duration_merge_samples) {
+				// 현재 구간의 끝점을 업데이트
+				currentSegment.end = speeches[i].end;
+			} else {
+				// 차이가 threshold(duration_merge_samples) 이상이면 현재 구간을 저장하고 새로운 구간 시작
+				mergedSpeeches.push_back(currentSegment);
+				currentSegment = speeches[i];
+			}
+		}
+
+		// 마지막 구간 추가
+		mergedSpeeches.push_back(currentSegment);
+
+		return mergedSpeeches;
+	}
+
+	}

examples/cpp_libtorch/silero_torch.h

@@ -0,0 +1,79 @@
+//Author      : Nathan Lee
+//Created On  : 2024-11-18
+//Description : silero 5.1 system for torch-script(c++).
+//Version     : 1.0
+//Contact     : junghan4242@gmail.com
+
+#ifndef SILERO_TORCH_H
+#define SILERO_TORCH_H
+
+#include <string>
+#include <memory>
+#include <stdexcept>
+#include <iostream>
+#include <memory>
+#include <vector>
+#include <fstream>
+#include <chrono>
+
+#include <torch/torch.h>
+#include <torch/script.h>
+
+
+namespace silero{
+
+	struct SpeechSegment{
+		int start;
+		int end;
+	};
+
+	struct Interval {
+		float start;
+		float end;
+	};
+
+	class VadIterator{
+		public:
+
+			VadIterator(const std::string &model_path, float threshold = 0.5, int sample_rate = 16000, 
+				int window_size_ms = 32, int speech_pad_ms = 30, int min_silence_duration_ms = 100, 
+				int min_speech_duration_ms = 250, int max_duration_merge_ms = 300, bool print_as_samples = false);
+			~VadIterator(); 
+
+
+			void SpeechProbs(std::vector<float>& input_wav);
+			std::vector<silero::Interval> GetSpeechTimestamps();
+
+			float threshold;
+			int sample_rate;
+			int min_speech_duration_ms;
+			int max_duration_merge_ms;
+			bool print_as_samples;
+
+		private:
+			torch::jit::script::Module model;
+			std::vector<float> outputs_prob;
+			int min_silence_samples;
+			int min_speech_samples;
+			int speech_pad_samples;
+			int window_size_samples;
+			int duration_merge_samples;
+			int current_sample = 0;
+
+			int total_sample_size=0;
+
+			int min_silence_duration_ms;
+			int speech_pad_ms;
+			bool triggered = false;
+			int temp_end = 0;
+
+			void init_engine(int window_size_ms);
+			void init_torch_model(const std::string& model_path);
+			void reset_states();
+			std::vector<Interval> DoVad();
+			std::vector<Interval> mergeSpeeches(const std::vector<Interval>& speeches, int duration_merge_samples);
+
+	};
+
+}
+#endif // SILERO_TORCH_H

examples/cpp_libtorch/wav.h

@@ -0,0 +1,235 @@
+// Copyright (c) 2016 Personal (Binbin Zhang)
+//
+// 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.
+
+
+#ifndef FRONTEND_WAV_H_
+#define FRONTEND_WAV_H_
+
+#include <assert.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <string>
+
+// #include "utils/log.h"
+
+namespace wav {
+
+struct WavHeader {
+  char riff[4];  // "riff"
+  unsigned int size;
+  char wav[4];  // "WAVE"
+  char fmt[4];  // "fmt "
+  unsigned int fmt_size;
+  uint16_t format;
+  uint16_t channels;
+  unsigned int sample_rate;
+  unsigned int bytes_per_second;
+  uint16_t block_size;
+  uint16_t bit;
+  char data[4];  // "data"
+  unsigned int data_size;
+};
+
+class WavReader {
+ public:
+  WavReader() : data_(nullptr) {}
+  explicit WavReader(const std::string& filename) { Open(filename); }
+
+  bool Open(const std::string& filename) {
+    FILE* fp = fopen(filename.c_str(), "rb"); //文件读取
+    if (NULL == fp) {
+      std::cout << "Error in read " << filename;
+      return false;
+    }
+
+    WavHeader header;
+    fread(&header, 1, sizeof(header), fp);
+    if (header.fmt_size < 16) {
+      printf("WaveData: expect PCM format data "
+              "to have fmt chunk of at least size 16.\n");
+      return false;
+    } else if (header.fmt_size > 16) {
+      int offset = 44 - 8 + header.fmt_size - 16;
+      fseek(fp, offset, SEEK_SET);
+      fread(header.data, 8, sizeof(char), fp);
+    }
+    // check "riff" "WAVE" "fmt " "data"
+
+    // Skip any sub-chunks between "fmt" and "data".  Usually there will
+    // be a single "fact" sub chunk, but on Windows there can also be a
+    // "list" sub chunk.
+    while (0 != strncmp(header.data, "data", 4)) {
+      // We will just ignore the data in these chunks.
+      fseek(fp, header.data_size, SEEK_CUR);
+      // read next sub chunk
+      fread(header.data, 8, sizeof(char), fp);
+    }
+
+    if (header.data_size == 0) {
+        int offset = ftell(fp);
+        fseek(fp, 0, SEEK_END);
+        header.data_size = ftell(fp) - offset;
+        fseek(fp, offset, SEEK_SET);
+    }
+
+    num_channel_ = header.channels;
+    sample_rate_ = header.sample_rate;
+    bits_per_sample_ = header.bit;
+    int num_data = header.data_size / (bits_per_sample_ / 8);
+    data_ = new float[num_data]; // Create 1-dim array
+    num_samples_ = num_data / num_channel_;
+
+    std::cout << "num_channel_    :" << num_channel_ << std::endl;
+    std::cout << "sample_rate_    :" << sample_rate_ << std::endl;
+    std::cout << "bits_per_sample_:" << bits_per_sample_ << std::endl;
+    std::cout << "num_samples     :" << num_data << std::endl;
+    std::cout << "num_data_size   :" << header.data_size << std::endl;
+
+    switch (bits_per_sample_) {
+        case 8: {
+            char sample;
+            for (int i = 0; i < num_data; ++i) {
+                fread(&sample, 1, sizeof(char), fp);
+                data_[i] = static_cast<float>(sample) / 32768;
+            }
+            break;
+        }
+        case 16: {
+            int16_t sample;
+            for (int i = 0; i < num_data; ++i) {
+                fread(&sample, 1, sizeof(int16_t), fp);
+                data_[i] = static_cast<float>(sample) / 32768;
+            }
+            break;
+        }
+        case 32:
+        {
+            if (header.format == 1) //S32
+            {
+                int sample;
+                for (int i = 0; i < num_data; ++i) {
+                    fread(&sample, 1, sizeof(int), fp);
+                    data_[i] = static_cast<float>(sample) / 32768;
+                }
+            }
+            else if (header.format == 3) // IEEE-float
+            {
+                float sample;
+                for (int i = 0; i < num_data; ++i) {
+                    fread(&sample, 1, sizeof(float), fp);
+                    data_[i] = static_cast<float>(sample);
+                }
+            }
+            else {
+                printf("unsupported quantization bits\n");
+            }
+            break;
+        }
+        default:
+            printf("unsupported quantization bits\n");
+            break;
+    }
+
+    fclose(fp);
+    return true;
+  }
+
+  int num_channel() const { return num_channel_; }
+  int sample_rate() const { return sample_rate_; }
+  int bits_per_sample() const { return bits_per_sample_; }
+  int num_samples() const { return num_samples_; }
+
+  ~WavReader() {
+    delete[] data_;
+  }
+
+  const float* data() const { return data_; }
+
+ private:
+  int num_channel_;
+  int sample_rate_;
+  int bits_per_sample_;
+  int num_samples_;  // sample points per channel
+  float* data_;
+};
+
+class WavWriter {
+ public:
+  WavWriter(const float* data, int num_samples, int num_channel,
+            int sample_rate, int bits_per_sample)
+      : data_(data),
+        num_samples_(num_samples),
+        num_channel_(num_channel),
+        sample_rate_(sample_rate),
+        bits_per_sample_(bits_per_sample) {}
+
+  void Write(const std::string& filename) {
+    FILE* fp = fopen(filename.c_str(), "w");
+    // init char 'riff' 'WAVE' 'fmt ' 'data'
+    WavHeader header;
+    char wav_header[44] = {0x52, 0x49, 0x46, 0x46, 0x00, 0x00, 0x00, 0x00, 0x57,
+                           0x41, 0x56, 0x45, 0x66, 0x6d, 0x74, 0x20, 0x10, 0x00,
+                           0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+                           0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+                           0x64, 0x61, 0x74, 0x61, 0x00, 0x00, 0x00, 0x00};
+    memcpy(&header, wav_header, sizeof(header));
+    header.channels = num_channel_;
+    header.bit = bits_per_sample_;
+    header.sample_rate = sample_rate_;
+    header.data_size = num_samples_ * num_channel_ * (bits_per_sample_ / 8);
+    header.size = sizeof(header) - 8 + header.data_size;
+    header.bytes_per_second =
+        sample_rate_ * num_channel_ * (bits_per_sample_ / 8);
+    header.block_size = num_channel_ * (bits_per_sample_ / 8);
+
+    fwrite(&header, 1, sizeof(header), fp);
+
+    for (int i = 0; i < num_samples_; ++i) {
+      for (int j = 0; j < num_channel_; ++j) {
+        switch (bits_per_sample_) {
+          case 8: {
+            char sample = static_cast<char>(data_[i * num_channel_ + j]);
+            fwrite(&sample, 1, sizeof(sample), fp);
+            break;
+          }
+          case 16: {
+            int16_t sample = static_cast<int16_t>(data_[i * num_channel_ + j]);
+            fwrite(&sample, 1, sizeof(sample), fp);
+            break;
+          }
+          case 32: {
+            int sample = static_cast<int>(data_[i * num_channel_ + j]);
+            fwrite(&sample, 1, sizeof(sample), fp);
+            break;
+          }
+        }
+      }
+    }
+    fclose(fp);
+  }
+
+ private:
+  const float* data_;
+  int num_samples_;  // total float points in data_
+  int num_channel_;
+  int sample_rate_;
+  int bits_per_sample_;
+};
+
+}  // namespace wenet
+
+#endif  // FRONTEND_WAV_H_