cpp/sherpa_kws_ax.cpp

#include <algorithm>
#include <array>
#include <chrono>
#include <cmath>
#include <cstdint>
#include <cstdio>
#include <cstring>
#include <deque>
#include <fstream>
#include <limits>
#include <memory>
#include <queue>
#include <sstream>
#include <stdexcept>
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>

#include "ax_engine_api.h"
#include "ax_sys_api.h"
#include "kaldi-native-fbank/csrc/online-feature.h"
#include "src/engine_wrapper.hpp"
#include "src/wav_reader.hpp"

#ifndef AXERA_TARGET_NAME
#define AXERA_TARGET_NAME "AXERA"
#endif

namespace {
constexpr int kSampleRate = 16000;
constexpr int kFeatureDim = 80;
constexpr int kEncoderDim = 320;
constexpr int kVocabSize = 263;
constexpr int kBlankId = 0;
constexpr int kUnkId = 2;
constexpr int kContextSize = 2;
using Clock = std::chrono::steady_clock;

double ElapsedSeconds(Clock::time_point begin, Clock::time_point end) {
  return std::chrono::duration<double>(end - begin).count();
}

struct Args {
  std::string models_dir = "models";
  std::string tokens = "config/tokens.txt";
  std::string keywords = "config/keywords.txt";
  std::string initial_decoder = "config/sherpa_decoder_initial.bin";
  std::string audio = "audio/sherpa/zh_0.wav";
  int chunk_size = 8;
  float default_score = 1.0f;
  float default_threshold = 0.25f;
  int trailing_blanks = 1;
  int max_active_paths = 1;
};

void Usage(const char *program) {
  std::printf(
      "Usage: %s [--models-dir DIR] [--tokens FILE] [--keywords FILE] "
      "[--initial-decoder-output FILE] [--audio WAV] [--chunk-size 8|16] "
      "[--keywords-score VALUE] [--keywords-threshold VALUE] "
      "[--num-trailing-blanks N] [--max-active-paths N]\n",
      program);
}

Args ParseArgs(int argc, char **argv) {
  Args args;
  for (int i = 1; i < argc; ++i) {
    const std::string key = argv[i];
    auto value = [&]() -> std::string {
      if (++i >= argc) throw std::runtime_error("Missing value for " + key);
      return argv[i];
    };
    if (key == "--models-dir") {
      args.models_dir = value();
    } else if (key == "--tokens") {
      args.tokens = value();
    } else if (key == "--keywords") {
      args.keywords = value();
    } else if (key == "--initial-decoder-output") {
      args.initial_decoder = value();
    } else if (key == "--audio") {
      args.audio = value();
    } else if (key == "--chunk-size") {
      args.chunk_size = std::stoi(value());
    } else if (key == "--keywords-score") {
      args.default_score = std::stof(value());
    } else if (key == "--keywords-threshold") {
      args.default_threshold = std::stof(value());
    } else if (key == "--num-trailing-blanks") {
      args.trailing_blanks = std::stoi(value());
    } else if (key == "--max-active-paths") {
      args.max_active_paths = std::stoi(value());
    } else if (key == "-h" || key == "--help") {
      Usage(argv[0]);
      std::exit(0);
    } else {
      throw std::runtime_error("Unknown argument: " + key);
    }
  }
  if (args.chunk_size != 8 && args.chunk_size != 16) {
    throw std::runtime_error("--chunk-size must be 8 or 16");
  }
  if (args.max_active_paths < 1 || args.max_active_paths > 32) {
    throw std::runtime_error("--max-active-paths must be in [1, 32]");
  }
  return args;
}

std::string Join(const std::string &left, const std::string &right) {
  return left.empty() || left.back() == '/' ? left + right
                                             : left + "/" + right;
}

std::string ModelPath(const Args &args, const std::string &component) {
  return Join(args.models_dir,
              "sherpa__" + component + "-epoch-13-avg-2-chunk-" +
                  std::to_string(args.chunk_size) + "-left-64.axmodel");
}

class AxRuntime {
 public:
  AxRuntime() {
    if (AX_SYS_Init() != 0) throw std::runtime_error("AX_SYS_Init failed");
    sys_initialized_ = true;
    AX_ENGINE_NPU_ATTR_T attr{};
    if (AX_ENGINE_Init(&attr) != 0) {
      AX_SYS_Deinit();
      sys_initialized_ = false;
      throw std::runtime_error("AX_ENGINE_Init failed");
    }
    engine_initialized_ = true;
  }
  ~AxRuntime() {
    if (engine_initialized_) AX_ENGINE_Deinit();
    if (sys_initialized_) AX_SYS_Deinit();
  }

 private:
  bool sys_initialized_ = false;
  bool engine_initialized_ = false;
};

std::unordered_map<std::string, int32_t> LoadTokens(const std::string &path) {
  std::ifstream input(path);
  if (!input) throw std::runtime_error("Cannot open tokens: " + path);
  std::unordered_map<std::string, int32_t> result;
  std::string line;
  int line_number = 0;
  while (std::getline(input, line)) {
    ++line_number;
    const std::size_t split = line.find_last_of(' ');
    if (split == std::string::npos) {
      throw std::runtime_error("Invalid token line " +
                               std::to_string(line_number));
    }
    result[line.substr(0, split)] = std::stoi(line.substr(split + 1));
  }
  return result;
}

struct Keyword {
  std::vector<int32_t> tokens;
  std::string phrase;
  float score = 1.0f;
  float threshold = 0.25f;
};

std::vector<Keyword> LoadKeywords(
    const std::string &path,
    const std::unordered_map<std::string, int32_t> &token_table,
    float default_score, float default_threshold) {
  std::ifstream input(path);
  if (!input) throw std::runtime_error("Cannot open keywords: " + path);
  std::vector<Keyword> result;
  std::string line;
  int line_number = 0;
  while (std::getline(input, line)) {
    ++line_number;
    std::istringstream stream(line);
    std::string part;
    Keyword keyword;
    keyword.score = default_score;
    keyword.threshold = default_threshold;
    while (stream >> part) {
      if (part.front() == '@') {
        keyword.phrase = part.substr(1);
      } else if (part.front() == ':') {
        keyword.score = std::stof(part.substr(1));
      } else if (part.front() == '#') {
        keyword.threshold = std::stof(part.substr(1));
      } else {
        const auto it = token_table.find(part);
        if (it == token_table.end()) {
          throw std::runtime_error("Unknown keyword token at line " +
                                   std::to_string(line_number) + ": " + part);
        }
        keyword.tokens.push_back(it->second);
      }
    }
    if (!keyword.tokens.empty()) {
      if (keyword.phrase.empty()) keyword.phrase = line;
      result.push_back(std::move(keyword));
    }
  }
  if (result.empty()) throw std::runtime_error("No keywords found in " + path);
  return result;
}

struct ContextNode {
  int32_t token = -1;
  int level = 0;
  float token_score = 0.0f;
  float node_score = 0.0f;
  float output_score = 0.0f;
  bool is_end = false;
  std::string phrase;
  float threshold = 0.0f;
  std::unordered_map<int32_t, std::unique_ptr<ContextNode>> children;
  ContextNode *fail = nullptr;
  ContextNode *output = nullptr;
};

class ContextGraph {
 public:
  struct Transition {
    float score = 0.0f;
    ContextNode *state = nullptr;
    ContextNode *matched = nullptr;
  };

  explicit ContextGraph(const std::vector<Keyword> &keywords) {
    root_.fail = &root_;
    for (const Keyword &keyword : keywords) {
      ContextNode *node = &root_;
      for (std::size_t i = 0; i < keyword.tokens.size(); ++i) {
        const int32_t token = keyword.tokens[i];
        auto &child = node->children[token];
        if (!child) {
          child = std::make_unique<ContextNode>();
          child->token = token;
          child->level = static_cast<int>(i + 1);
          child->token_score = keyword.score;
          child->node_score = node->node_score + keyword.score;
          const bool is_end = i + 1 == keyword.tokens.size();
          child->output_score = is_end ? child->node_score : 0.0f;
          child->is_end = is_end;
        } else {
          child->token_score = std::max(child->token_score, keyword.score);
          child->node_score = node->node_score + child->token_score;
          child->is_end = child->is_end || i + 1 == keyword.tokens.size();
          child->output_score = child->is_end ? child->node_score : 0.0f;
        }
        node = child.get();
      }
      node->is_end = true;
      node->phrase = keyword.phrase;
      node->threshold = keyword.threshold;
    }
    FillFailureLinks();
  }

  ContextNode *Root() { return &root_; }

  Transition Forward(ContextNode *state, int32_t token) {
    ContextNode *node = nullptr;
    float score = 0.0f;
    const auto direct = state->children.find(token);
    if (direct != state->children.end()) {
      node = direct->second.get();
      score = node->token_score;
    } else {
      node = state->fail;
      while (node->children.count(token) == 0) {
        node = node->fail;
        if (node->token == -1) break;
      }
      const auto fallback = node->children.find(token);
      if (fallback != node->children.end()) node = fallback->second.get();
      score = node->node_score - state->node_score;
    }
    ContextNode *matched = node->is_end ? node : node->output;
    return {score + node->output_score, node, matched};
  }

  ContextNode *Matched(ContextNode *state) {
    return state->is_end ? state : state->output;
  }

 private:
  void FillFailureLinks() {
    std::queue<ContextNode *> queue;
    for (auto &entry : root_.children) {
      entry.second->fail = &root_;
      queue.push(entry.second.get());
    }
    while (!queue.empty()) {
      ContextNode *current = queue.front();
      queue.pop();
      for (auto &entry : current->children) {
        const int32_t token = entry.first;
        ContextNode *child = entry.second.get();
        ContextNode *failure = current->fail;
        while (failure != &root_ && failure->children.count(token) == 0) {
          failure = failure->fail;
        }
        const auto it = failure->children.find(token);
        child->fail = (it != failure->children.end() && it->second.get() != child)
                          ? it->second.get()
                          : &root_;
        ContextNode *output = child->fail;
        while (output != &root_ && !output->is_end) output = output->fail;
        child->output = output->is_end ? output : nullptr;
        if (child->output) child->output_score += child->output->output_score;
        queue.push(child);
      }
    }
  }

  ContextNode root_;
};

std::array<float, kEncoderDim> LoadInitialDecoder(const std::string &path) {
  std::ifstream input(path, std::ios::binary);
  if (!input) {
    throw std::runtime_error("Cannot open initial decoder output: " + path);
  }
  char magic[8]{};
  input.read(magic, 8);
  uint32_t version = 0;
  uint32_t count = 0;
  input.read(reinterpret_cast<char *>(&version), sizeof(version));
  input.read(reinterpret_cast<char *>(&count), sizeof(count));
  if (std::memcmp(magic, "SHDEC1", 6) != 0 || version != 1 ||
      count != kEncoderDim) {
    throw std::runtime_error("Invalid initial decoder output file");
  }
  std::array<float, kEncoderDim> result{};
  input.read(reinterpret_cast<char *>(result.data()),
             result.size() * sizeof(float));
  if (!input) throw std::runtime_error("Truncated initial decoder output");
  return result;
}

std::vector<float> ComputeFbank(const PcmWav &wav) {
  if (wav.sample_rate != kSampleRate) {
    throw std::runtime_error("Input WAV must use 16 kHz sample rate");
  }
  knf::FbankOptions options;
  options.frame_opts.samp_freq = kSampleRate;
  options.frame_opts.dither = 0.0f;
  options.frame_opts.frame_length_ms = 25.0f;
  options.frame_opts.frame_shift_ms = 10.0f;
  options.frame_opts.snip_edges = false;
  options.frame_opts.window_type = "povey";
  options.mel_opts.num_bins = kFeatureDim;
  options.mel_opts.low_freq = 20.0f;
  options.mel_opts.high_freq = -400.0f;
  options.energy_floor = 0.0f;
  knf::OnlineFbank fbank(options);
  std::vector<float> waveform(wav.samples.size());
  for (std::size_t i = 0; i < wav.samples.size(); ++i) {
    waveform[i] = static_cast<float>(wav.samples[i]) / 32768.0f;
  }
  fbank.AcceptWaveform(kSampleRate, waveform.data(),
                       static_cast<int32_t>(waveform.size()));
  std::vector<float> tail(static_cast<std::size_t>(0.8f * kSampleRate));
  fbank.AcceptWaveform(kSampleRate, tail.data(),
                       static_cast<int32_t>(tail.size()));
  fbank.InputFinished();
  const int frames = fbank.NumFramesReady();
  std::vector<float> result(static_cast<std::size_t>(frames) * kFeatureDim);
  for (int i = 0; i < frames; ++i) {
    std::memcpy(result.data() + static_cast<std::size_t>(i) * kFeatureDim,
                fbank.GetFrame(i), kFeatureDim * sizeof(float));
  }
  return result;
}

class KeywordDecoder {
 public:
  KeywordDecoder(EngineWrapper *decoder, EngineWrapper *joiner,
                 ContextGraph *graph,
                 const std::array<float, kEncoderDim> &initial_decoder,
                 int trailing_blanks, int max_active_paths)
      : decoder_(decoder),
        joiner_(joiner),
        graph_(graph),
        initial_decoder_(initial_decoder),
        required_trailing_blanks_(trailing_blanks),
        max_active_paths_(max_active_paths) {
    Reset();
  }

  void Reset() {
    hypotheses_.clear();
    Hypothesis initial;
    initial.history = {-1, kBlankId};
    initial.context_state = graph_->Root();
    hypotheses_.push_back(std::move(initial));
  }

  int trailing_blanks() const { return BestHypothesis().trailing_blanks; }

  std::string DecodeFrame(const float *encoder_frame) {
    std::vector<Candidate> candidates;
    candidates.reserve(hypotheses_.size() * kVocabSize);
    for (std::size_t hyp_index = 0; hyp_index < hypotheses_.size();
         ++hyp_index) {
      const Hypothesis &hypothesis = hypotheses_[hyp_index];
      const std::array<float, kEncoderDim> decoder_output =
          DecoderOutput(hypothesis);
      if (joiner_->SetInputByName("encoder_out", encoder_frame,
                                  kEncoderDim * sizeof(float)) != 0 ||
          joiner_->SetInputByName("decoder_out", decoder_output.data(),
                                  kEncoderDim * sizeof(float)) != 0 ||
          joiner_->RunSync() != 0) {
        throw std::runtime_error("Joiner inference failed");
      }
      std::array<float, kVocabSize> logits{};
      if (joiner_->GetOutputByName("logit", logits.data(),
                                   logits.size() * sizeof(float)) != 0) {
        throw std::runtime_error("Failed to read joiner output");
      }
      const float maximum = *std::max_element(logits.begin(), logits.end());
      double sum = 0.0;
      for (float value : logits) sum += std::exp(value - maximum);
      const double log_normalizer = maximum + std::log(sum);
      for (int32_t token = 0; token < kVocabSize; ++token) {
        Candidate candidate;
        candidate.hypothesis = hyp_index;
        candidate.token = token;
        candidate.acoustic_probability =
            static_cast<float>(std::exp(logits[token] - log_normalizer));
        candidate.selection_score =
            hypothesis.log_probability + logits[token] - log_normalizer;
        candidate.score = candidate.selection_score;
        candidate.context_state = hypothesis.context_state;
        if (token != kBlankId && token != kUnkId) {
          const auto transition = graph_->Forward(hypothesis.context_state, token);
          candidate.score += transition.score;
          candidate.context_state = transition.state;
        }
        candidates.push_back(candidate);
      }
    }

    const std::size_t keep = std::min<std::size_t>(
        static_cast<std::size_t>(max_active_paths_), candidates.size());
    std::partial_sort(candidates.begin(), candidates.begin() + keep,
                      candidates.end(),
                      [](const Candidate &left, const Candidate &right) {
                        return left.selection_score > right.selection_score;
                      });
    std::unordered_map<std::string, Hypothesis> merged;
    for (std::size_t i = 0; i < keep; ++i) {
      const Candidate &candidate = candidates[i];
      Hypothesis next = hypotheses_[candidate.hypothesis];
      next.log_probability = candidate.score;
      if (candidate.token != kBlankId && candidate.token != kUnkId) {
        next.history.push_back(candidate.token);
        next.probabilities.push_back(candidate.acoustic_probability);
        next.trailing_blanks = 0;
        next.context_state = candidate.context_state;
        if (next.context_state == graph_->Root()) {
          next.history = {-1, kBlankId};
          next.probabilities.clear();
        }
      } else {
        ++next.trailing_blanks;
      }
      const std::string key = HistoryKey(next.history);
      const auto existing = merged.find(key);
      if (existing == merged.end()) {
        merged.emplace(key, std::move(next));
      } else {
        existing->second.log_probability =
            LogAdd(existing->second.log_probability, next.log_probability);
      }
    }
    hypotheses_.clear();
    hypotheses_.reserve(merged.size());
    for (auto &entry : merged) hypotheses_.push_back(std::move(entry.second));

    const Hypothesis &best = BestHypothesis();
    ContextNode *matched = graph_->Matched(best.context_state);
    if (!matched || best.trailing_blanks <= required_trailing_blanks_ ||
        best.probabilities.size() < static_cast<std::size_t>(matched->level)) {
      return {};
    }
    float acoustic_score = 0.0f;
    const std::size_t begin = best.probabilities.size() - matched->level;
    for (std::size_t i = begin; i < best.probabilities.size(); ++i) {
      acoustic_score += best.probabilities[i];
    }
    acoustic_score /= matched->level;
    if (acoustic_score < matched->threshold) return {};
    const std::string phrase = matched->phrase;
    Reset();
    return phrase;
  }

 private:
  struct Hypothesis {
    std::vector<int32_t> history;
    std::vector<float> probabilities;
    ContextNode *context_state = nullptr;
    int trailing_blanks = 0;
    double log_probability = 0.0;
  };

  struct Candidate {
    std::size_t hypothesis = 0;
    int32_t token = 0;
    float acoustic_probability = 0.0f;
    double selection_score = 0.0;
    double score = 0.0;
    ContextNode *context_state = nullptr;
  };

  static double LogAdd(double left, double right) {
    const double maximum = std::max(left, right);
    return maximum + std::log(std::exp(left - maximum) +
                              std::exp(right - maximum));
  }

  static std::string HistoryKey(const std::vector<int32_t> &history) {
    std::string result;
    for (int32_t token : history) {
      if (!result.empty()) result.push_back('-');
      result += std::to_string(token);
    }
    return result;
  }

  const Hypothesis &BestHypothesis() const {
    if (hypotheses_.empty()) throw std::runtime_error("No active hypotheses");
    return *std::max_element(
        hypotheses_.begin(), hypotheses_.end(),
        [](const Hypothesis &left, const Hypothesis &right) {
          return left.log_probability < right.log_probability;
        });
  }

  std::array<float, kEncoderDim> DecoderOutput(
      const Hypothesis &hypothesis) {
    if (hypothesis.history[hypothesis.history.size() - 2] < 0 ||
        hypothesis.history.back() < 0) {
      return initial_decoder_;
    }
    const std::pair<int32_t, int32_t> key{
        hypothesis.history[hypothesis.history.size() - 2],
        hypothesis.history.back()};
    const auto cached = cache_.find(key);
    if (cached != cache_.end()) return cached->second;
    const std::array<int32_t, kContextSize> decoder_input{key.first, key.second};
    if (decoder_->SetInputByName("y", decoder_input.data(),
                                 decoder_input.size() * sizeof(int32_t)) != 0 ||
        decoder_->RunSync() != 0) {
      throw std::runtime_error("Decoder inference failed");
    }
    std::array<float, kEncoderDim> output{};
    if (decoder_->GetOutputByName("decoder_out", output.data(),
                                  output.size() * sizeof(float)) != 0) {
      throw std::runtime_error("Failed to read decoder output");
    }
    cache_.emplace(key, output);
    return output;
  }

  struct PairHash {
    std::size_t operator()(const std::pair<int32_t, int32_t> &value) const {
      return (static_cast<std::size_t>(static_cast<uint32_t>(value.first))
              << 32) ^
             static_cast<uint32_t>(value.second);
    }
  };

  EngineWrapper *decoder_;
  EngineWrapper *joiner_;
  ContextGraph *graph_;
  std::array<float, kEncoderDim> initial_decoder_{};
  int required_trailing_blanks_ = 1;
  int max_active_paths_ = 1;
  std::vector<Hypothesis> hypotheses_;
  std::unordered_map<std::pair<int32_t, int32_t>,
                     std::array<float, kEncoderDim>, PairHash>
      cache_;
};

void ResetEncoderStates(EngineWrapper *encoder) {
  for (std::size_t i = 1; i < encoder->InputCount(); ++i) {
    if (encoder->ZeroInputByName(encoder->InputName(i)) != 0) {
      throw std::runtime_error("Failed to reset encoder state: " +
                               encoder->InputName(i));
    }
  }
}

void UpdateEncoderStates(EngineWrapper *encoder) {
  for (std::size_t i = 1; i < encoder->InputCount(); ++i) {
    const std::string &input_name = encoder->InputName(i);
    if (encoder->CopyOutputToInputByName("new_" + input_name, input_name) != 0) {
      throw std::runtime_error("Failed to update encoder state: " + input_name);
    }
  }
}

void Run(const Args &args) {
  const PcmWav wav = ReadPcmWav(args.audio);
  const double audio_seconds =
      static_cast<double>(wav.samples.size()) / kSampleRate;
  if (audio_seconds <= 0.0) {
    throw std::runtime_error("Input WAV contains no samples");
  }
  const auto feature_begin = Clock::now();
  const std::vector<float> features = ComputeFbank(wav);
  const double feature_seconds = ElapsedSeconds(feature_begin, Clock::now());
  const int feature_frames = static_cast<int>(features.size() / kFeatureDim);
  const auto token_table = LoadTokens(args.tokens);
  const auto keywords = LoadKeywords(args.keywords, token_table,
                                     args.default_score,
                                     args.default_threshold);
  ContextGraph graph(keywords);
  const auto initial_decoder = LoadInitialDecoder(args.initial_decoder);

  const auto model_load_begin = Clock::now();
  AxRuntime runtime;
  EngineWrapper encoder;
  EngineWrapper decoder;
  EngineWrapper joiner;
  if (encoder.Init(ModelPath(args, "encoder")) != 0 ||
      decoder.Init(ModelPath(args, "decoder")) != 0 ||
      joiner.Init(ModelPath(args, "joiner")) != 0) {
    throw std::runtime_error("Failed to load Sherpa KWS axmodels");
  }
  const double model_load_seconds =
      ElapsedSeconds(model_load_begin, Clock::now());
  KeywordDecoder keyword_decoder(&decoder, &joiner, &graph, initial_decoder,
                                 args.trailing_blanks, args.max_active_paths);
  ResetEncoderStates(&encoder);

  const int input_frames = args.chunk_size == 8 ? 29 : 45;
  const int output_frames = args.chunk_size == 8 ? 4 : 8;
  const int chunk_shift = args.chunk_size * 2;
  if (encoder.GetInputSizeByName("x") !=
      input_frames * kFeatureDim * static_cast<int>(sizeof(float))) {
    throw std::runtime_error("Encoder input shape does not match chunk size");
  }
  std::vector<std::string> detections;
  int decode_calls = 0;
  const auto inference_begin = Clock::now();
  for (int start = 0; start + input_frames < feature_frames;
       start += chunk_shift) {
    if (keyword_decoder.trailing_blanks() * 0.04f > 1.5f) {
      ResetEncoderStates(&encoder);
      keyword_decoder.Reset();
    }
    const float *input =
        features.data() + static_cast<std::size_t>(start) * kFeatureDim;
    if (encoder.SetInputByName(
            "x", input,
            input_frames * kFeatureDim * sizeof(float)) != 0 ||
        encoder.RunSync() != 0) {
      throw std::runtime_error("Encoder inference failed");
    }
    std::vector<float> encoder_output(
        static_cast<std::size_t>(output_frames) * kEncoderDim);
    if (encoder.GetOutputByName("encoder_out", encoder_output.data(),
                                encoder_output.size() * sizeof(float)) != 0) {
      throw std::runtime_error("Failed to read encoder output");
    }
    UpdateEncoderStates(&encoder);
    ++decode_calls;
    bool found = false;
    for (int frame = 0; frame < output_frames; ++frame) {
      const std::string phrase = keyword_decoder.DecodeFrame(
          encoder_output.data() + frame * kEncoderDim);
      if (!phrase.empty()) {
        detections.push_back(phrase);
        found = true;
      }
    }
    if (found) {
      ResetEncoderStates(&encoder);
      keyword_decoder.Reset();
    }
  }
  const double inference_seconds =
      ElapsedSeconds(inference_begin, Clock::now());
  const double processing_seconds = feature_seconds + inference_seconds;
  const double rtf = processing_seconds / audio_seconds;

  std::printf("\nSherpa KWS C++ inference complete\n");
  std::printf(
      "target: %s\naudio: %s\nchunk_size: %d\nmax_active_paths: %d\n"
      "feature_frames: %d\n",
      AXERA_TARGET_NAME, args.audio.c_str(), args.chunk_size,
      args.max_active_paths,
      feature_frames);
  std::printf("decode_calls: %d\ndetections:", decode_calls);
  if (detections.empty()) {
    std::printf(" []\n");
  } else {
    std::printf("\n");
    for (const std::string &phrase : detections) {
      std::printf("  WAKEUP %s\n", phrase.c_str());
    }
  }
  std::printf("audio_seconds: %.6f\n", audio_seconds);
  std::printf("feature_seconds: %.6f\n", feature_seconds);
  std::printf("model_load_seconds: %.6f\n", model_load_seconds);
  std::printf("inference_seconds: %.6f\n", inference_seconds);
  std::printf("processing_seconds: %.6f\n", processing_seconds);
  std::printf("rtf: %.6f\n", rtf);
}
}  // namespace

int main(int argc, char **argv) {
  try {
    const Args args = ParseArgs(argc, argv);
    Run(args);
    return 0;
  } catch (const std::exception &error) {
    std::fprintf(stderr, "ERROR: %s\n", error.what());
    return 1;
  }
}