cpp/inference/audio_encoder_lib.cpp

#include "audio_encoder_lib.h"

#include <iostream>
#include <fstream>
#include <cmath>
#include <numeric>
#include <algorithm>
#include <cstring> // For memcpy

// Include specific ONNX Runtime headers for implementation
#include <onnxruntime_cxx_api.h>

// Include specific Eigen headers for implementation
#include <Eigen/Dense>

// Include specific KissFFT headers for implementation
#include <kiss_fft.h>
#include <kiss_fftr.h>

// Define M_PI if it's not already defined
#ifndef M_PI
#define M_PI 3.14159265358979323846
#endif

// --- Global parameters for feature extraction (matching Python script) ---
// These are constants derived from the Python preprocessing script and are
// internal to the feature extraction logic.
namespace { // Anonymous namespace for internal linkage
    const float PREEMPHASIS_COEFF = 0.97f;
    const int N_FFT = 512;       // FFT size
    const int WIN_LENGTH = 400;  // Window length (samples)
    const int HOP_LENGTH = 160;  // Hop length (samples)
    const int N_MELS = 80;       // Number of Mel filterbank channels
    const int TARGET_SAMPLE_RATE = 16000; // Target sample rate for feature extraction
}

// --- Implementation of AudioInferenceEngine methods ---

AudioInferenceEngine::AudioInferenceEngine(const std::string& modelPath) {
    // 1. Initialize ONNX Runtime Environment
    env_ = std::make_unique<Ort::Env>(ORT_LOGGING_LEVEL_WARNING, "AudioInferenceEngine");

    // 2. Configure Session Options
    Ort::SessionOptions session_options;
    session_options.SetIntraOpNumThreads(0);
    session_options.SetGraphOptimizationLevel(ORT_ENABLE_EXTENDED);

    // 3. Create ONNX Runtime Session
    session_ = std::make_unique<Ort::Session>(*env_, modelPath.c_str(), session_options);

    // 4. Initialize Allocator
    allocator_ = std::make_unique<Ort::AllocatorWithDefaultOptions>();

    // 5. Get Input and Output Node Names
    // It's crucial to allocate these names using the allocator and store them
    // as C-style strings for Ort::Session::Run.
    size_t numInputNodes = session_->GetInputCount();
    if (numInputNodes == 0) {
        throw Ort::Exception("ONNX model has no input nodes.", ORT_FAIL);
    }
    input_node_names_.resize(numInputNodes);
    for (size_t i = 0; i < numInputNodes; ++i) {
        input_node_names_[i] = session_->GetInputNameAllocated(i, *allocator_).release(); // release() to manage lifetime
    }

    size_t numOutputNodes = session_->GetOutputCount();
    if (numOutputNodes == 0) {
        throw Ort::Exception("ONNX model has no output nodes.", ORT_FAIL);
    }
    output_node_names_.resize(numOutputNodes);
    for (size_t i = 0; i < numOutputNodes; ++i) {
        output_node_names_[i] = session_->GetOutputNameAllocated(i, *allocator_).release(); // release() to manage lifetime
    }

    // 6. Precompute Mel filterbank
    // The Python example uses fmax=16000//2-80-230.
    float mel_fmax = static_cast<float>(TARGET_SAMPLE_RATE) / 2.0f - 80.0f - 230.0f;
    mel_filterbank_ = speechlibMel(TARGET_SAMPLE_RATE, N_FFT, N_MELS, 0.0f, mel_fmax);

    if (mel_filterbank_.rows() == 0 || mel_filterbank_.cols() == 0) {
        throw std::runtime_error("Failed to create Mel filterbank during initialization.");
    }

    std::cout << "AudioInferenceEngine initialized successfully with model: " << modelPath << std::endl;
}

AudioInferenceEngine::~AudioInferenceEngine() {
    // Release allocated names
    for (const char* name : input_node_names_) {
        allocator_->Free(const_cast<void*>(reinterpret_cast<const void*>(name)));
    }
    for (const char* name : output_node_names_) {
        allocator_->Free(const_cast<void*>(reinterpret_cast<const void*>(name)));
    }
    // unique_ptr automatically handles deletion of env_ and session_
}

/**
 * @brief Private helper: Loads audio data from a WAV file.
 */
std::vector<float> AudioInferenceEngine::loadWavToFloatArray(const std::string& filename, int& actual_sample_rate) {
    std::ifstream file(filename, std::ios::binary);
    if (!file.is_open()) {
        std::cerr << "Error: Could not open WAV file: " << filename << std::endl;
        return {};
    }

    WavHeader header;
    file.read(reinterpret_cast<char*>(&header), sizeof(WavHeader));

    if (std::string(header.riff_id, 4) != "RIFF" ||
        std::string(header.wave_id, 4) != "WAVE" ||
        std::string(header.fmt_id, 4) != "fmt ") {
        std::cerr << "Error: Invalid WAV header (RIFF, WAVE, or fmt chunk missing/invalid)." << std::endl;
        file.close();
        return {};
    }

    if (header.audio_format != 1) { // 1 = PCM
        std::cerr << "Error: Only PCM audio format (1) is supported. Found: " << header.audio_format << std::endl;
        file.close();
        return {};
    }

    if (header.bits_per_sample != 16) {
        std::cerr << "Error: Only 16-bit PCM is supported. Found: " << header.bits_per_sample << " bits per sample." << std::endl;
        file.close();
        return {};
    }

    actual_sample_rate = header.sample_rate;

    WavDataChunk data_chunk;
    bool data_chunk_found = false;
    while (!file.eof()) {
        file.read(reinterpret_cast<char*>(&data_chunk.data_id), 4);
        file.read(reinterpret_cast<char*>(&data_chunk.data_size), 4);

        if (std::string(data_chunk.data_id, 4) == "data") {
            data_chunk_found = true;
            break;
        } else {
            file.seekg(data_chunk.data_size, std::ios::cur);
        }
    }

    if (!data_chunk_found) {
        std::cerr << "Error: 'data' chunk not found in WAV file." << std::endl;
        file.close();
        return {};
    }

    std::vector<float> audioData;
    int16_t sample_buffer;
    long num_samples_to_read = data_chunk.data_size / sizeof(int16_t);

    for (long i = 0; i < num_samples_to_read; ++i) {
        file.read(reinterpret_cast<char*>(&sample_buffer), sizeof(int16_t));
        float normalized_sample = static_cast<float>(sample_buffer) / 32768.0f;

        if (header.num_channels == 1) {
            audioData.push_back(normalized_sample);
        } else if (header.num_channels == 2) {
            int16_t right_sample;
            if (file.read(reinterpret_cast<char*>(&right_sample), sizeof(int16_t))) {
                float normalized_right_sample = static_cast<float>(right_sample) / 32768.0f;
                audioData.push_back((normalized_sample + normalized_right_sample) / 2.0f);
                i++;
            } else {
                std::cerr << "Warning: Unexpected end of file while reading stereo data." << std::endl;
                break;
            }
        } else {
            std::cerr << "Error: Unsupported number of channels: " << header.num_channels << std::endl;
            file.close();
            return {};
        }
    }

    file.close();
    return audioData;
}

/**
 * @brief Private helper: Generates a Hamming window.
 */
std::vector<float> AudioInferenceEngine::generateHammingWindow(int window_length) {
    std::vector<float> window(window_length);
    for (int i = 0; i < window_length; ++i) {
        window[i] = 0.54f - 0.46f * std::cos(2 * M_PI * i / static_cast<float>(window_length - 1));
    }
    return window;
}

/**
 * @brief Private helper: Extracts spectrogram features.
 */
Eigen::MatrixXf AudioInferenceEngine::extractSpectrogram(const std::vector<float>& wav, int fs) {
    int n_batch = (wav.size() - WIN_LENGTH) / HOP_LENGTH + 1;
    if (n_batch <= 0) {
        return Eigen::MatrixXf(0, N_FFT / 2 + 1);
    }

    std::vector<float> fft_window = generateHammingWindow(WIN_LENGTH);

    kiss_fftr_cfg fft_cfg = kiss_fftr_alloc(N_FFT, 0 /* is_inverse_fft */, nullptr, nullptr);
    if (!fft_cfg) {
        std::cerr << "Error: Failed to allocate KissFFT configuration." << std::endl;
        return Eigen::MatrixXf(0, N_FFT / 2 + 1);
    }

    Eigen::MatrixXf spec_matrix(n_batch, N_FFT / 2 + 1);

    std::vector<float> frame_buffer(WIN_LENGTH);
    kiss_fft_scalar fft_input[N_FFT];
    kiss_fft_cpx fft_output[N_FFT / 2 + 1];

    for (int i = 0; i < n_batch; ++i) {
        int start_idx = i * HOP_LENGTH;

        for (int j = 0; j < WIN_LENGTH; ++j) {
            frame_buffer[j] = wav[start_idx + j];
        }

        // Apply pre-emphasis and scale by 32768
        if (WIN_LENGTH > 0) {
            if (WIN_LENGTH > 1) {
                // Corrected pre-emphasis to match Python's np.roll and then overwrite first element
                // The first element of the frame is pre-emphasized against the second element.
                fft_input[0] = (frame_buffer[0] - PREEMPHASIS_COEFF * frame_buffer[1]) * 32768.0f;
                for (int j = 1; j < WIN_LENGTH; ++j) {
                    fft_input[j] = (frame_buffer[j] - PREEMPHASIS_COEFF * frame_buffer[j - 1]) * 32768.0f;
                }
            } else { // WIN_LENGTH == 1
                fft_input[0] = frame_buffer[0] * 32768.0f;
            }
        }
        for (int j = WIN_LENGTH; j < N_FFT; ++j) {
            fft_input[j] = 0.0f;
        }

        for (int j = 0; j < WIN_LENGTH; ++j) {
            fft_input[j] *= fft_window[j];
        }

        kiss_fftr(fft_cfg, fft_input, fft_output);

        for (int j = 0; j <= N_FFT / 2; ++j) {
            spec_matrix(i, j) = std::sqrt(fft_output[j].r * fft_output[j].r + fft_output[j].i * fft_output[j].i);
        }
    }

    kiss_fftr_free(fft_cfg);
    return spec_matrix;
}

/**
 * @brief Private helper: Creates a Mel filter-bank matrix.
 */
Eigen::MatrixXf AudioInferenceEngine::speechlibMel(int sample_rate, int n_fft, int n_mels, float fmin, float fmax) {
    int bank_width = n_fft / 2 + 1;
    if (fmax == 0.0f) fmax = sample_rate / 2.0f;
    if (fmin == 0.0f) fmin = 0.0f;

    auto mel = [](float f) { return 1127.0f * std::log(1.0f + f / 700.0f); };
    auto bin2mel = [&](int fft_bin) { return 1127.0f * std::log(1.0f + static_cast<float>(fft_bin) * sample_rate / (static_cast<float>(n_fft) * 700.0f)); };
    auto f2bin = [&](float f) { return static_cast<int>((f * n_fft / sample_rate) + 0.5f); };

    int klo = f2bin(fmin) + 1;
    int khi = f2bin(fmax);
    khi = std::max(khi, klo);

    float mlo = mel(fmin);
    float mhi = mel(fmax);
    
    std::vector<float> m_centers(n_mels + 2);
    float ms = (mhi - mlo) / (n_mels + 1);
    for (int i = 0; i < n_mels + 2; ++i) {
        m_centers[i] = mlo + i * ms;
    }

    Eigen::MatrixXf matrix = Eigen::MatrixXf::Zero(n_mels, bank_width);

    for (int m = 0; m < n_mels; ++m) {
        float left = m_centers[m];
        float center = m_centers[m + 1];
        float right = m_centers[m + 2];
        for (int fft_bin = klo; fft_bin < bank_width; ++fft_bin) {
            float mbin = bin2mel(fft_bin);
            if (left < mbin && mbin < right) {
                matrix(m, fft_bin) = 1.0f - std::abs(center - mbin) / ms;
            }
        }
    }
    return matrix;
}

/**
 * @brief Public method: Preprocesses an audio WAV file.
 */
Eigen::MatrixXf AudioInferenceEngine::preprocessAudio(const std::string& wavFilePath) {
    int actual_wav_sample_rate = 0;
    std::vector<float> audioWav = loadWavToFloatArray(wavFilePath, actual_wav_sample_rate);

    if (audioWav.empty()) {
        std::cerr << "Failed to load audio data from " << wavFilePath << "." << std::endl;
        return Eigen::MatrixXf(0, N_MELS);
    }

    if (actual_wav_sample_rate != TARGET_SAMPLE_RATE) {
        std::cerr << "Warning: WAV file sample rate (" << actual_wav_sample_rate
                  << " Hz) does not match the target sample rate for feature extraction ("
                  << TARGET_SAMPLE_RATE << " Hz)." << std::endl;
        std::cerr << "This example does NOT include resampling. Features will be extracted at "
                  << TARGET_SAMPLE_RATE << " Hz, which might lead to incorrect results if the WAV file's sample rate is different." << std::endl;
    }

    Eigen::MatrixXf spec = extractSpectrogram(audioWav, TARGET_SAMPLE_RATE);
    if (spec.rows() == 0) {
        std::cerr << "Error: Spectrogram extraction failed." << std::endl;
        return Eigen::MatrixXf(0, N_MELS);
    }

    Eigen::MatrixXf spec_power = spec.array().square();
    Eigen::MatrixXf fbank_power = spec_power * mel_filterbank_.transpose(); // Transpose mel_filterbank_ for correct multiplication

    fbank_power = fbank_power.array().max(1.0f);
    Eigen::MatrixXf log_fbank = fbank_power.array().log();

    return log_fbank;
}

/**
 * @brief Public method: Runs inference on the loaded ONNX model.
 */
std::vector<float> AudioInferenceEngine::runInference(const Eigen::MatrixXf& features) {
    if (features.rows() == 0 || features.cols() == 0) {
        std::cerr << "Error: Input features are empty for inference." << std::endl;
        return {};
    }

    // Prepare Input Tensor Shape: [batch, frames, feature_size]
    std::vector<int64_t> inputTensorShape = {1, features.rows(), features.cols()};

    // Flatten Eigen::MatrixXf into std::vector<float> in row-major order
    std::vector<float> inputTensorData(features.rows() * features.cols());
    for (int r = 0; r < features.rows(); ++r) {
        for (int c = 0; c < features.cols(); ++c) {
            inputTensorData[r * features.cols() + c] = features(r, c);
        }
    }

    Ort::MemoryInfo memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
    Ort::Value inputTensor = Ort::Value::CreateTensor<float>(memory_info, inputTensorData.data(), inputTensorData.size(),
                                                              inputTensorShape.data(), inputTensorShape.size());

    if (!inputTensor.IsTensor()) {
        std::cerr << "Error: Created input tensor is not valid!" << std::endl;
        return {};
    }

    // Run Inference
    std::vector<Ort::Value> outputTensors = session_->Run(Ort::RunOptions{nullptr},
                                                          input_node_names_.data(), &inputTensor, 1,
                                                          output_node_names_.data(), output_node_names_.size());

    if (outputTensors.empty() || !outputTensors[0].IsTensor()) {
        std::cerr << "Error: No valid output tensors received from the model." << std::endl;
        return {};
    }

    // Copy output data
    float* outputData = outputTensors[0].GetTensorMutableData<float>();
    Ort::TensorTypeAndShapeInfo outputShapeInfo = outputTensors[0].GetTensorTypeAndShapeInfo();
    size_t outputSize = outputShapeInfo.GetElementCount();

    std::vector<float> result(outputData, outputData + outputSize);
    return result;
}

std::vector<Ort::Value> AudioInferenceEngine::runInference_tensor(const Ort::Value& inputTensor) {
    // Run Inference
    std::vector<Ort::Value> outputTensors = session_->Run(Ort::RunOptions{nullptr},
                                                          input_node_names_.data(), &inputTensor, 1,
                                                          output_node_names_.data(), output_node_names_.size());

    return outputTensors;
}