micro_wake_word.cpp

Go to the documentation of this file.

#include "micro_wake_word.h"
 
#ifdef USE_ESP32
 
#include "esphome/core/application.h"
#include "esphome/core/hal.h"
#include "esphome/core/helpers.h"
#include "esphome/core/log.h"
 
#include "esphome/components/audio/audio_transfer_buffer.h"
 
#include <algorithm>
 
#ifdef USE_OTA
#include "esphome/components/ota/ota_backend.h"
#endif
 
namespace esphome::micro_wake_word {
 
// Config validation checks the YAML, this checks the build: a unicore sdkconfig reached any other way
static_assert(MICRO_WAKE_WORD_TASK_CORE == tskNO_AFFINITY || MICRO_WAKE_WORD_TASK_CORE < portNUM_PROCESSORS,
              "micro_wake_word task_core is not a core of this build");
 
ESPHOME_LOG_TAG(TAG, "micro_wake_word");
 
static const ssize_t DETECTION_QUEUE_LENGTH = 5;
 
static const size_t DATA_TIMEOUT_MS = 50;
 
#ifdef CONFIG_IDF_TARGET_ESP32P4
// ESP32-P4 PIE-optimized esp-nn kernels (e.g. depthwise_conv_s8_ch1_pie) require
// significantly more stack than other variants, causing stack protection faults at 3072.
static const uint32_t INFERENCE_TASK_STACK_SIZE = 8192;
#else
static const uint32_t INFERENCE_TASK_STACK_SIZE = 3072;
#endif
static const UBaseType_t INFERENCE_TASK_PRIORITY = 3;
 
enum EventGroupBits : uint32_t {
  COMMAND_STOP = (1 << 0),               // Signals the inference task should stop
  COMMAND_RESET_RING_BUFFER = (1 << 1),  // Signals the inference task to discard buffered audio
  COMMAND_PAUSE_MODELS = (1 << 2),       // Asks the inference task to pause at a safe point so the model lists can be
                                         // mutated from the main loop
 
  TASK_STARTING = (1 << 3),
  TASK_RUNNING = (1 << 4),
  TASK_STOPPING = (1 << 5),
  TASK_STOPPED = (1 << 6),
 
  MODELS_PAUSED = (1 << 7),          // Inference task acknowledges it is paused and holds no iterators
  COMMAND_RESUME_MODELS = (1 << 8),  // Main loop signals the inference task it may resume iterating
 
  ERROR_MEMORY = (1 << 9),
  ERROR_INFERENCE = (1 << 10),
 
  WARNING_FULL_RING_BUFFER = (1 << 13),
  WARNING_MODELS_RESUME_TIMEOUT = (1 << 14),  // The paused inference task gave up waiting to be released
 
  ERROR_BITS = ERROR_MEMORY | ERROR_INFERENCE,
  ALL_BITS = 0xfffff,  // 24 total bits available in an event group
};
 
// How long the main loop waits for the inference task to acknowledge a pause request before giving up.
// The task checks for the command at the top of its loop, which runs at least every DATA_TIMEOUT_MS.
static const uint32_t MODELS_PAUSE_TIMEOUT_MS = 500;
// How long the paused inference task waits to be resumed before rechecking on its own. Only reached if
// the main loop abandoned the handshake (e.g. it timed out first), so recovery just needs to be bounded.
static const uint32_t MODELS_RESUME_TIMEOUT_MS = 1000;
 
float MicroWakeWord::get_setup_priority() const { return setup_priority::AFTER_CONNECTION; }
 
static const LogString *micro_wake_word_state_to_string(State state) {
  switch (state) {
    case State::STARTING:
      return LOG_STR("STARTING");
    case State::DETECTING_WAKE_WORD:
      return LOG_STR("DETECTING_WAKE_WORD");
    case State::STOPPING:
      return LOG_STR("STOPPING");
    case State::STOPPED:
      return LOG_STR("STOPPED");
    default:
      return LOG_STR("UNKNOWN");
  }
}
 
void MicroWakeWord::dump_config() {
  ESP_LOGCONFIG(TAG, "microWakeWord:");
  // tskNO_AFFINITY is a cast expression, so this cannot be a preprocessor test; the compiler folds it
  if (MICRO_WAKE_WORD_TASK_CORE == tskNO_AFFINITY) {
    ESP_LOGCONFIG(TAG, "  Inference task core: any");
  } else {
    ESP_LOGCONFIG(TAG, "  Inference task core: %d", MICRO_WAKE_WORD_TASK_CORE);
  }
  ESP_LOGCONFIG(TAG, "  models:");
  for (auto &model : this->wake_word_models_) {
    model->log_model_config();
  }
#ifdef USE_MICRO_WAKE_WORD_VAD
  this->vad_model_->log_model_config();
#endif
}
 
void MicroWakeWord::setup() {
  this->frontend_config_.window.size_ms = FEATURE_DURATION_MS;
  this->frontend_config_.window.step_size_ms = this->features_step_size_;
  this->frontend_config_.filterbank.num_channels = PREPROCESSOR_FEATURE_SIZE;
  this->frontend_config_.filterbank.lower_band_limit = FILTERBANK_LOWER_BAND_LIMIT;
  this->frontend_config_.filterbank.upper_band_limit = FILTERBANK_UPPER_BAND_LIMIT;
  this->frontend_config_.noise_reduction.smoothing_bits = NOISE_REDUCTION_SMOOTHING_BITS;
  this->frontend_config_.noise_reduction.even_smoothing = NOISE_REDUCTION_EVEN_SMOOTHING;
  this->frontend_config_.noise_reduction.odd_smoothing = NOISE_REDUCTION_ODD_SMOOTHING;
  this->frontend_config_.noise_reduction.min_signal_remaining = NOISE_REDUCTION_MIN_SIGNAL_REMAINING;
  this->frontend_config_.pcan_gain_control.enable_pcan = PCAN_GAIN_CONTROL_ENABLE_PCAN;
  this->frontend_config_.pcan_gain_control.strength = PCAN_GAIN_CONTROL_STRENGTH;
  this->frontend_config_.pcan_gain_control.offset = PCAN_GAIN_CONTROL_OFFSET;
  this->frontend_config_.pcan_gain_control.gain_bits = PCAN_GAIN_CONTROL_GAIN_BITS;
  this->frontend_config_.log_scale.enable_log = LOG_SCALE_ENABLE_LOG;
  this->frontend_config_.log_scale.scale_shift = LOG_SCALE_SCALE_SHIFT;
 
  this->event_group_ = xEventGroupCreate();
  if (this->event_group_ == nullptr) {
    ESP_LOGE(TAG, "Failed to create event group");
    this->mark_failed();
    return;
  }
 
  this->detection_queue_ = xQueueCreate(DETECTION_QUEUE_LENGTH, sizeof(DetectionEvent));
  if (this->detection_queue_ == nullptr) {
    ESP_LOGE(TAG, "Failed to create detection event queue");
    this->mark_failed();
    return;
  }
 
  this->microphone_source_->add_data_callback([this](const std::vector<uint8_t> &data) {
    if (this->state_ == State::STOPPED) {
      return;
    }
    std::shared_ptr<ring_buffer::RingBuffer> temp_ring_buffer = this->ring_buffer_.lock();
    if (temp_ring_buffer != nullptr) {
      // Producer-only write: never touches consumer state. If the buffer is full, ask the inference task
      // to drain it - reset() is a consumer operation and must run on the inference task's thread.
      // Disable partial writes so audio chunks are either fully accepted or rejected and handled below.
      if (temp_ring_buffer->write_without_replacement(data.data(), data.size(), 0, false) == 0) {
        xEventGroupSetBits(this->event_group_,
                           EventGroupBits::WARNING_FULL_RING_BUFFER | EventGroupBits::COMMAND_RESET_RING_BUFFER);
      }
    }
  });
 
#ifdef USE_OTA_STATE_LISTENER
  ota::get_global_ota_callback()->add_global_state_listener(this);
#endif
}
 
#ifdef USE_OTA_STATE_LISTENER
void MicroWakeWord::on_ota_global_state(ota::OTAState state, float progress, uint8_t error, ota::OTAComponent *comp) {
  if (state == ota::OTA_STARTED) {
    this->suspend_task_();
  } else if (state == ota::OTA_ERROR) {
    this->resume_task_();
  }
}
#endif
 
void MicroWakeWord::inference_task(void *params) {
  MicroWakeWord *this_mww = (MicroWakeWord *) params;
 
  xEventGroupSetBits(this_mww->event_group_, EventGroupBits::TASK_STARTING);
 
  {  // Ensures any C++ objects fall out of scope to deallocate before deleting the task
 
    const auto &stream_info = this_mww->microphone_source_->get_audio_stream_info();
    const size_t bytes_per_frame = stream_info.frames_to_bytes(1);
    const size_t max_fill_bytes = stream_info.ms_to_bytes(this_mww->features_step_size_);
    std::unique_ptr<audio::RingBufferAudioSource> audio_source;
    int8_t features_buffer[PREPROCESSOR_FEATURE_SIZE];
 
    if (!(xEventGroupGetBits(this_mww->event_group_) & ERROR_BITS)) {
      // Round ring buffer size down to a frame multiple so the wrap boundary never splits an int16 sample.
      const size_t ring_buffer_size =
          (stream_info.ms_to_bytes(this_mww->ring_buffer_duration_ms_) / bytes_per_frame) * bytes_per_frame;
      std::shared_ptr<ring_buffer::RingBuffer> temp_ring_buffer = ring_buffer::RingBuffer::create(ring_buffer_size);
      if (temp_ring_buffer == nullptr) {
        xEventGroupSetBits(this_mww->event_group_, EventGroupBits::ERROR_MEMORY);
      } else {
        audio_source = audio::RingBufferAudioSource::create(temp_ring_buffer, max_fill_bytes,
                                                            static_cast<uint8_t>(bytes_per_frame));
        if (audio_source == nullptr) {
          xEventGroupSetBits(this_mww->event_group_, EventGroupBits::ERROR_MEMORY);
        } else {
          this_mww->ring_buffer_ = temp_ring_buffer;
        }
      }
    }
 
    if (!(xEventGroupGetBits(this_mww->event_group_) & ERROR_BITS)) {
      this_mww->microphone_source_->start();
      xEventGroupSetBits(this_mww->event_group_, EventGroupBits::TASK_RUNNING);
 
      while (!(xEventGroupGetBits(this_mww->event_group_) & (COMMAND_STOP | ERROR_BITS))) {
        if (xEventGroupGetBits(this_mww->event_group_) & EventGroupBits::COMMAND_PAUSE_MODELS) {
          // Safe point: no iterators into wake_word_models_ are held here. Acknowledge the pause and wait for the
          // main loop to finish mutating the model lists before resuming.
          xEventGroupSetBits(this_mww->event_group_, EventGroupBits::MODELS_PAUSED);
          EventBits_t resume_bits = xEventGroupWaitBits(this_mww->event_group_, EventGroupBits::COMMAND_RESUME_MODELS,
                                                        pdTRUE, pdTRUE, pdMS_TO_TICKS(MODELS_RESUME_TIMEOUT_MS));
          if (!(resume_bits & EventGroupBits::COMMAND_RESUME_MODELS)) {
            // Nobody released us, so the main loop abandoned the handshake and did not mutate the lists.
            // Rechecking the pause command below is safe, but the wait cost a second of detection, so report it.
            xEventGroupSetBits(this_mww->event_group_, EventGroupBits::WARNING_MODELS_RESUME_TIMEOUT);
          }
          continue;
        }
 
        if (xEventGroupGetBits(this_mww->event_group_) & EventGroupBits::COMMAND_RESET_RING_BUFFER) {
          // Producer asked us to drain; run the consumer-side reset from this thread.
          audio_source->clear_buffered_data();
          xEventGroupClearBits(this_mww->event_group_, EventGroupBits::COMMAND_RESET_RING_BUFFER);
        }
 
        audio_source->fill(pdMS_TO_TICKS(DATA_TIMEOUT_MS));
 
        // The frontend buffers samples internally and only emits a feature once it has a full window, so we can
        // hand it whatever the source exposes. The frontend consumes at least one sample per call, so available()
        // strictly decreases and this loop always terminates.
        while (audio_source->available() >= sizeof(int16_t)) {
          const size_t samples_available = audio_source->available() / sizeof(int16_t);
          const int16_t *audio_data = reinterpret_cast<const int16_t *>(audio_source->data());
 
          size_t processed_samples = 0;
          const bool feature_generated =
              this_mww->generate_features_(audio_data, samples_available, features_buffer, &processed_samples);
          audio_source->consume(processed_samples * sizeof(int16_t));
 
          if (feature_generated) {
            if (!this_mww->update_model_probabilities_(features_buffer)) {
              xEventGroupSetBits(this_mww->event_group_, EventGroupBits::ERROR_INFERENCE);
              break;
            }
 
            // Process each model's probabilities and possibly send a Detection Event to the queue
            this_mww->process_probabilities_();
          }
        }
      }
    }
  }
 
  xEventGroupSetBits(this_mww->event_group_, EventGroupBits::TASK_STOPPING);
 
  this_mww->unload_models_();
  this_mww->microphone_source_->stop();
  FrontendFreeStateContents(&this_mww->frontend_state_);
 
  xEventGroupSetBits(this_mww->event_group_, EventGroupBits::TASK_STOPPED);
  vTaskSuspend(nullptr);  // Suspend this task indefinitely until the loop method deletes it
}
 
std::vector<WakeWordModel *> MicroWakeWord::get_wake_words() {
  std::vector<WakeWordModel *> external_wake_word_models;
  for (auto *model : this->wake_word_models_) {
    if (!model->get_internal_only()) {
      external_wake_word_models.push_back(model);
    }
  }
  return external_wake_word_models;
}
 
void MicroWakeWord::add_wake_word_model(WakeWordModel *model) { this->wake_word_models_.push_back(model); }
 
bool MicroWakeWord::try_lock_models_() {
  // When the inference task isn't running it holds no iterators into wake_word_models_, so the lists can be
  // mutated without a handshake. The main loop is the only caller, so this state cannot change between here
  // and the matching unlock_models_() call.
  if (!this->inference_task_.is_created() || this->state_ == State::STOPPED) {
    return true;
  }
 
  // The task is running and iterates wake_word_models_. Ask it to pause at a safe point before we mutate.
  // Clear any stale acknowledgement from an abandoned handshake first.
  xEventGroupClearBits(this->event_group_, EventGroupBits::MODELS_PAUSED);
  xEventGroupSetBits(this->event_group_, EventGroupBits::COMMAND_PAUSE_MODELS);
 
  EventBits_t bits = xEventGroupWaitBits(this->event_group_, EventGroupBits::MODELS_PAUSED, pdFALSE, pdTRUE,
                                         pdMS_TO_TICKS(MODELS_PAUSE_TIMEOUT_MS));
 
  if (!(bits & EventGroupBits::MODELS_PAUSED)) {
    // The task never acknowledged (e.g. it is busy stopping). Withdraw the request and refuse to mutate a
    // list it might be iterating.
    xEventGroupClearBits(this->event_group_, EventGroupBits::COMMAND_PAUSE_MODELS);
    return false;
  }
  return true;
}
 
void MicroWakeWord::unlock_models_() {
  if (!this->inference_task_.is_created() || this->state_ == State::STOPPED) {
    return;  // Nothing was paused
  }
  xEventGroupClearBits(this->event_group_, EventGroupBits::MODELS_PAUSED | EventGroupBits::COMMAND_PAUSE_MODELS);
  xEventGroupSetBits(this->event_group_, EventGroupBits::COMMAND_RESUME_MODELS);
}
 
bool MicroWakeWord::add_runtime_model(std::unique_ptr<WakeWordModel> model) {
  if (!model) {
    ESP_LOGE(TAG, "Cannot add null runtime model");
    return false;
  }
 
  const std::string model_id = model->get_id();
 
  // A model without usable data can never load, so keep it out of the lists entirely. Otherwise it would be
  // advertised to Home Assistant as selectable and the inference task would silently disable it again every
  // time it was enabled.
  if (!model->has_model_data()) {
    ESP_LOGE(TAG, "Runtime model '%s' has no valid data", model_id.c_str());
    return false;
  }
 
  // Reject a duplicate id against every model (compiled or runtime). The inference task only ever reads
  // wake_word_models_, so scanning it here (on the main loop) needs no synchronization.
  for (auto *existing : this->wake_word_models_) {
    if (existing->get_id() == model_id) {
      ESP_LOGW(TAG, "Wake word model '%s' already exists", model_id.c_str());
      return false;
    }
  }
 
  if (!this->try_lock_models_()) {
    ESP_LOGE(TAG, "Timed out pausing inference task; not adding runtime model '%s'", model_id.c_str());
    return false;
  }
 
  this->wake_word_models_.push_back(model.get());
  this->runtime_models_.push_back(std::move(model));
 
  this->unlock_models_();
  ESP_LOGD(TAG, "Added runtime model '%s'", model_id.c_str());
  return true;
}
 
bool MicroWakeWord::remove_runtime_model(const std::string &model_id) {
  // Only runtime-downloaded models can be removed; compiled-in models never appear in runtime_models_.
  auto runtime_it =
      std::find_if(this->runtime_models_.begin(), this->runtime_models_.end(),
                   [&model_id](const std::unique_ptr<WakeWordModel> &m) { return m->get_id() == model_id; });
  if (runtime_it == this->runtime_models_.end()) {
    return false;
  }
 
  if (!this->try_lock_models_()) {
    ESP_LOGE(TAG, "Timed out pausing inference task; not removing runtime model '%s'", model_id.c_str());
    return false;
  }
 
  WakeWordModel *raw = runtime_it->get();
  auto models_it = std::find(this->wake_word_models_.begin(), this->wake_word_models_.end(), raw);
  if (models_it != this->wake_word_models_.end()) {
    this->wake_word_models_.erase(models_it);
  }
 
  // Queued detection events hold a pointer into the model being destroyed, so drop them. The inference task
  // is parked, so no new events can be queued concurrently. Losing an undelivered detection from another
  // model is acceptable for this rare operation.
  xQueueReset(this->detection_queue_);
 
  // Free the interpreter and arenas (safe: the task is parked, not mid-inference), then destroy the model.
  // Its ModelData releases the PSRAM model buffer once the last shared_ptr reference drops.
  raw->unload_model();
  this->runtime_models_.erase(runtime_it);
 
  this->unlock_models_();
  ESP_LOGI(TAG, "Removed runtime model '%s'", model_id.c_str());
  return true;
}
 
std::vector<std::string> MicroWakeWord::get_runtime_model_ids() {
  std::vector<std::string> ids;
  ids.reserve(this->runtime_models_.size());
  for (const auto &model : this->runtime_models_) {
    ids.push_back(model->get_id());
  }
  return ids;
}
 
WakeWordModel *MicroWakeWord::get_model_by_id(const std::string &model_id) {
  for (auto *model : this->wake_word_models_) {
    if (model->get_id() == model_id) {
      return model;
    }
  }
  return nullptr;
}
 
#ifdef USE_MICRO_WAKE_WORD_VAD
void MicroWakeWord::add_vad_model(const uint8_t *model_start, size_t model_size, uint8_t probability_cutoff,
                                  size_t sliding_window_size, size_t tensor_arena_size) {
  this->vad_model_ =
      make_unique<VADModel>(model_start, model_size, probability_cutoff, sliding_window_size, tensor_arena_size);
}
#endif
 
void MicroWakeWord::suspend_task_() {
  if (this->inference_task_.is_created()) {
    vTaskSuspend(this->inference_task_.get_handle());
  }
}
 
void MicroWakeWord::resume_task_() {
  if (this->inference_task_.is_created()) {
    vTaskResume(this->inference_task_.get_handle());
  }
}
 
void MicroWakeWord::loop() {
  uint32_t event_group_bits = xEventGroupGetBits(this->event_group_);
 
  if (event_group_bits & EventGroupBits::ERROR_MEMORY) {
    xEventGroupClearBits(this->event_group_, EventGroupBits::ERROR_MEMORY);
    ESP_LOGE(TAG, "Encountered an error allocating buffers");
  }
 
  if (event_group_bits & EventGroupBits::ERROR_INFERENCE) {
    xEventGroupClearBits(this->event_group_, EventGroupBits::ERROR_INFERENCE);
    ESP_LOGE(TAG, "Encountered an error while performing an inference");
  }
 
  if (event_group_bits & EventGroupBits::WARNING_FULL_RING_BUFFER) {
    xEventGroupClearBits(this->event_group_, EventGroupBits::WARNING_FULL_RING_BUFFER);
    ESP_LOGW(TAG, "Not enough free bytes in ring buffer to store incoming audio data. Resetting the ring buffer. Wake "
                  "word detection accuracy will temporarily be reduced.");
  }
 
  if (event_group_bits & EventGroupBits::WARNING_MODELS_RESUME_TIMEOUT) {
    xEventGroupClearBits(this->event_group_, EventGroupBits::WARNING_MODELS_RESUME_TIMEOUT);
    ESP_LOGW(TAG, "Inference task paused for %" PRIu32 " ms without being released, so it resumed on its own",
             MODELS_RESUME_TIMEOUT_MS);
  }
 
  if (event_group_bits & EventGroupBits::TASK_STARTING) {
    ESP_LOGD(TAG, "Inference task has started, attempting to allocate memory for buffers");
    xEventGroupClearBits(this->event_group_, EventGroupBits::TASK_STARTING);
  }
 
  if (event_group_bits & EventGroupBits::TASK_RUNNING) {
    ESP_LOGD(TAG, "Inference task is running");
 
    xEventGroupClearBits(this->event_group_, EventGroupBits::TASK_RUNNING);
    this->set_state_(State::DETECTING_WAKE_WORD);
  }
 
  if (event_group_bits & EventGroupBits::TASK_STOPPING) {
    ESP_LOGD(TAG, "Inference task is stopping, deallocating buffers");
    xEventGroupClearBits(this->event_group_, EventGroupBits::TASK_STOPPING);
  }
 
  // Retries on a subsequent loop if the task is still running on the other core
  if ((event_group_bits & EventGroupBits::TASK_STOPPED) && this->inference_task_.deallocate()) {
    ESP_LOGD(TAG, "Inference task is finished, freeing task resources");
    xEventGroupClearBits(this->event_group_, ALL_BITS);
    xQueueReset(this->detection_queue_);
    this->set_state_(State::STOPPED);
  }
 
  if ((this->pending_start_) && (this->state_ == State::STOPPED)) {
    this->set_state_(State::STARTING);
    this->pending_start_ = false;
  }
 
  if ((this->pending_stop_) && (this->state_ == State::DETECTING_WAKE_WORD)) {
    this->set_state_(State::STOPPING);
    this->pending_stop_ = false;
  }
 
  switch (this->state_) {
    case State::STARTING:
      if (!this->inference_task_.is_created() && !this->status_has_error()) {
        // Setup preprocesor feature generator. If done in the task, it would lock the task to its initial core, as it
        // uses floating point operations.
        if (!FrontendPopulateState(&this->frontend_config_, &this->frontend_state_,
                                   this->microphone_source_->get_audio_stream_info().get_sample_rate())) {
          this->status_momentary_error(1000);
          return;
        }
 
        if (!this->inference_task_.create(MicroWakeWord::inference_task, "mww", INFERENCE_TASK_STACK_SIZE,
                                          (void *) this, INFERENCE_TASK_PRIORITY, this->task_stack_in_psram_,
                                          MICRO_WAKE_WORD_TASK_CORE)) {
          FrontendFreeStateContents(&this->frontend_state_);  // Deallocate frontend state
          this->status_momentary_error(1000);
        }
      }
      break;
    case State::DETECTING_WAKE_WORD: {
      DetectionEvent detection_event;
      while (xQueueReceive(this->detection_queue_, &detection_event, 0)) {
        if (detection_event.blocked_by_vad) {
          ESP_LOGD(TAG, "Wake word model predicts '%s', but VAD model doesn't.", detection_event.wake_word->c_str());
        } else {
          constexpr float uint8_to_float_divisor =
              255.0f;  // Converting a quantized uint8 probability to floating point
          ESP_LOGD(TAG, "Detected '%s' with sliding average probability is %.2f and max probability is %.2f",
                   detection_event.wake_word->c_str(), (detection_event.average_probability / uint8_to_float_divisor),
                   (detection_event.max_probability / uint8_to_float_divisor));
          this->wake_word_detected_trigger_.trigger(*detection_event.wake_word);
          if (this->stop_after_detection_) {
            this->stop();
          }
        }
      }
      break;
    }
    case State::STOPPING:
      xEventGroupSetBits(this->event_group_, EventGroupBits::COMMAND_STOP);
      break;
    case State::STOPPED:
      break;
  }
}
 
void MicroWakeWord::start() {
  if (!this->is_ready()) {
    ESP_LOGW(TAG, "Wake word detection can't start as the component hasn't been setup yet");
    return;
  }
 
  if (this->is_failed()) {
    ESP_LOGW(TAG, "Wake word component is marked as failed. Please check setup logs");
    return;
  }
 
  if (this->is_running()) {
    ESP_LOGW(TAG, "Wake word detection is already running");
    return;
  }
 
  ESP_LOGD(TAG, "Starting wake word detection");
 
  this->pending_start_ = true;
  this->pending_stop_ = false;
}
 
void MicroWakeWord::stop() {
  if (this->state_ == STOPPED)
    return;
 
  ESP_LOGD(TAG, "Stopping wake word detection");
 
  this->pending_start_ = false;
  this->pending_stop_ = true;
}
 
void MicroWakeWord::set_state_(State state) {
  if (this->state_ != state) {
    ESP_LOGD(TAG, "State changed from %s to %s", LOG_STR_ARG(micro_wake_word_state_to_string(this->state_)),
             LOG_STR_ARG(micro_wake_word_state_to_string(state)));
    this->state_ = state;
  }
}
 
bool MicroWakeWord::generate_features_(const int16_t *audio_buffer, size_t samples_available,
                                       int8_t features_buffer[PREPROCESSOR_FEATURE_SIZE], size_t *processed_samples) {
  *processed_samples = 0;
  struct FrontendOutput frontend_output =
      FrontendProcessSamples(&this->frontend_state_, audio_buffer, samples_available, processed_samples);
 
  if (frontend_output.size == 0) {
    return false;
  }
 
  for (size_t i = 0; i < frontend_output.size; ++i) {
    // These scaling values are set to match the TFLite audio frontend int8 output.
    // The feature pipeline outputs 16-bit signed integers in roughly a 0 to 670
    // range. In training, these are then arbitrarily divided by 25.6 to get
    // float values in the rough range of 0.0 to 26.0. This scaling is performed
    // for historical reasons, to match up with the output of other feature
    // generators.
    // The process is then further complicated when we quantize the model. This
    // means we have to scale the 0.0 to 26.0 real values to the -128 (INT8_MIN)
    // to 127 (INT8_MAX) signed integer numbers.
    // All this means that to get matching values from our integer feature
    // output into the tensor input, we have to perform:
    // input = (((feature / 25.6) / 26.0) * 256) - 128
    // To simplify this and perform it in 32-bit integer math, we rearrange to:
    // input = (feature * 256) / (25.6 * 26.0) - 128
    constexpr int32_t value_scale = 256;
    constexpr int32_t value_div = 666;  // 666 = 25.6 * 26.0 after rounding
    int32_t value = ((frontend_output.values[i] * value_scale) + (value_div / 2)) / value_div;
 
    value += INT8_MIN;  // Adds a -128; i.e., subtracts 128
    features_buffer[i] = static_cast<int8_t>(clamp<int32_t>(value, INT8_MIN, INT8_MAX));
  }
 
  return true;
}
 
void MicroWakeWord::process_probabilities_() {
#ifdef USE_MICRO_WAKE_WORD_VAD
  DetectionEvent vad_state = this->vad_model_->determine_detected();
 
  this->vad_state_ = vad_state.detected;  // atomic write, so thread safe
#endif
 
  for (auto &model : this->wake_word_models_) {
    if (model->get_unprocessed_probability_status()) {
      // Only detect wake words if there is a new probability since the last check
      DetectionEvent wake_word_state = model->determine_detected();
      if (wake_word_state.detected) {
#ifdef USE_MICRO_WAKE_WORD_VAD
        if (vad_state.detected) {
#endif
          xQueueSend(this->detection_queue_, &wake_word_state, portMAX_DELAY);
 
          // Wake main loop immediately to process wake word detection
          App.wake_loop_threadsafe();
 
          model->reset_probabilities();
#ifdef USE_MICRO_WAKE_WORD_VAD
        } else {
          wake_word_state.blocked_by_vad = true;
          xQueueSend(this->detection_queue_, &wake_word_state, portMAX_DELAY);
        }
#endif
      }
    }
  }
}
 
void MicroWakeWord::unload_models_() {
  for (auto &model : this->wake_word_models_) {
    model->unload_model();
  }
#ifdef USE_MICRO_WAKE_WORD_VAD
  this->vad_model_->unload_model();
#endif
}
 
bool MicroWakeWord::update_model_probabilities_(const int8_t audio_features[PREPROCESSOR_FEATURE_SIZE]) {
  bool success = true;
 
  for (auto &model : this->wake_word_models_) {
    // Perform inference
    success = success & model->perform_streaming_inference(audio_features);
  }
#ifdef USE_MICRO_WAKE_WORD_VAD
  success = success & this->vad_model_->perform_streaming_inference(audio_features);
#endif
 
  return success;
}
 
}  // namespace esphome::micro_wake_word
 
#endif  // USE_ESP32