Fix computing features for CED audio tagging models. (#1341)

See also
https://github.com/RicherMans/CED/blob/main/onnx_inference_with_kaldi.py

@@ -8,6 +8,7 @@
 #include <cassert>
 #include <cassert>
 #include <cmath>
 #include <cmath>
 #include <iomanip>
 #include <iomanip>
+#include <limits>
 #include <utility>
 #include <utility>
 
 
 #include "kaldi-native-fbank/csrc/online-feature.h"
 #include "kaldi-native-fbank/csrc/online-feature.h"
@@ -110,7 +111,7 @@ class OfflineStream::Impl {
     config_.sampling_rate = opts_.frame_opts.samp_freq;
     config_.sampling_rate = opts_.frame_opts.samp_freq;
   }
   }
 
 
-  explicit Impl(CEDTag /*tag*/) {
+  explicit Impl(CEDTag /*tag*/) : is_ced_(true) {
     // see
     // see
     // https://github.com/RicherMans/CED/blob/main/onnx_inference_with_kaldi.py
     // https://github.com/RicherMans/CED/blob/main/onnx_inference_with_kaldi.py
 
 
@@ -123,7 +124,9 @@ class OfflineStream::Impl {
 
 
     opts_.frame_opts.samp_freq = 16000;  // fixed to 16000
     opts_.frame_opts.samp_freq = 16000;  // fixed to 16000
     opts_.mel_opts.num_bins = 64;
     opts_.mel_opts.num_bins = 64;
+    opts_.mel_opts.low_freq = 0;
     opts_.mel_opts.high_freq = 8000;
     opts_.mel_opts.high_freq = 8000;
+    opts_.use_log_fbank = false;
 
 
     config_.sampling_rate = opts_.frame_opts.samp_freq;
     config_.sampling_rate = opts_.frame_opts.samp_freq;
 
 
@@ -216,6 +219,10 @@ class OfflineStream::Impl {
 
 
     NemoNormalizeFeatures(features.data(), n, feature_dim);
     NemoNormalizeFeatures(features.data(), n, feature_dim);
 
 
+    if (is_ced_) {
+      AmplitudeToDB(features.data(), features.size());
+    }
+
     return features;
     return features;
   }
   }
 
 
@@ -226,6 +233,32 @@ class OfflineStream::Impl {
   const ContextGraphPtr &GetContextGraph() const { return context_graph_; }
   const ContextGraphPtr &GetContextGraph() const { return context_graph_; }
 
 
  private:
  private:
+  // see
+  // https://github.com/pytorch/audio/blob/main/src/torchaudio/functional/functional.py#L359
+  void AmplitudeToDB(float *p, int32_t n) const {
+    float multiplier = 10;
+    float top_db = 120;
+    float amin = 1e-10;
+
+    float max_x = std::numeric_limits<float>::min();
+
+    for (int32_t i = 0; i != n; ++i) {
+      float x = p[i];
+      x = (x > amin) ? x : amin;
+      x = std::log10f(x) * multiplier;
+
+      max_x = (x > max_x) ? x : max_x;
+      p[i] = x;
+    }
+
+    float d = max_x - top_db;
+    for (int32_t i = 0; i != n; ++i) {
+      float x = p[i];
+      x = (x > d) ? x : d;
+      p[i] = x;
+    }
+  }
+
   void NemoNormalizeFeatures(float *p, int32_t num_frames,
   void NemoNormalizeFeatures(float *p, int32_t num_frames,
                              int32_t feature_dim) const {
                              int32_t feature_dim) const {
     if (config_.nemo_normalize_type.empty()) {
     if (config_.nemo_normalize_type.empty()) {
@@ -266,6 +299,7 @@ class OfflineStream::Impl {
   knf::MfccOptions mfcc_opts_;
   knf::MfccOptions mfcc_opts_;
   OfflineRecognitionResult r_;
   OfflineRecognitionResult r_;
   ContextGraphPtr context_graph_;
   ContextGraphPtr context_graph_;
+  bool is_ced_ = false;
 };
 };
 
 
 OfflineStream::OfflineStream(const FeatureExtractorConfig &config /*= {}*/,
 OfflineStream::OfflineStream(const FeatureExtractorConfig &config /*= {}*/,