[src] Clarify feature extraction options; make features independent of BaseFloat

src/feat/feature-fbank.cc

@@ -83,7 +83,7 @@ void FbankComputer::Compute(BaseFloat signal_log_energy,
   // Compute energy after window function (not the raw one).
   if (opts_.use_energy && !opts_.raw_energy)
     signal_log_energy = Log(std::max(VecVec(*signal_frame, *signal_frame),
-                                     std::numeric_limits<BaseFloat>::min()));
+                                     std::numeric_limits<float>::min()));
 
   if (srfft_ != NULL)  // Compute FFT using split-radix algorithm.
     srfft_->Compute(signal_frame->Data(), true);
@@ -108,7 +108,7 @@ void FbankComputer::Compute(BaseFloat signal_log_energy,
   mel_banks.Compute(power_spectrum, &mel_energies);
   if (opts_.use_log_fbank) {
     // Avoid log of zero (which should be prevented anyway by dithering).
-    mel_energies.ApplyFloor(std::numeric_limits<BaseFloat>::epsilon());
+    mel_energies.ApplyFloor(std::numeric_limits<float>::epsilon());
     mel_energies.ApplyLog();  // take the log.
   }
 

src/feat/feature-fbank.h

@@ -53,7 +53,7 @@ struct FbankOptions {
                  // this seems to be common for 16khz-sampled data,
                  // but for 8khz-sampled data, 15 may be better.
                  use_energy(false),
-                 energy_floor(0.0),  // not in log scale: a small value e.g. 1.0e-10
+                 energy_floor(0.0),
                  raw_energy(true),
                  htk_compat(false),
                  use_log_fbank(true),
@@ -65,7 +65,9 @@ struct FbankOptions {
     opts->Register("use-energy", &use_energy,
                    "Add an extra dimension with energy to the FBANK output.");
     opts->Register("energy-floor", &energy_floor,
-                   "Floor on energy (absolute, not relative) in FBANK computation");
+                   "Floor on energy (absolute, not relative) in MFCC computation. "
+                   "Only makes a difference if --use-energy=true; only necessary if "
+                   "--dither=0.0.  Suggested values: 0.1 or 1.0");
     opts->Register("raw-energy", &raw_energy,
                    "If true, compute energy before preemphasis and windowing");
     opts->Register("htk-compat", &htk_compat, "If true, put energy last.  "

src/feat/feature-mfcc.cc

@@ -36,7 +36,7 @@ void MfccComputer::Compute(BaseFloat signal_log_energy,
 
   if (opts_.use_energy && !opts_.raw_energy)
     signal_log_energy = Log(std::max(VecVec(*signal_frame, *signal_frame),
-                                     std::numeric_limits<BaseFloat>::min()));
+                                     std::numeric_limits<float>::min()));
 
   if (srfft_ != NULL)  // Compute FFT using the split-radix algorithm.
     srfft_->Compute(signal_frame->Data(), true);
@@ -51,7 +51,7 @@ void MfccComputer::Compute(BaseFloat signal_log_energy,
   mel_banks.Compute(power_spectrum, &mel_energies_);
 
   // avoid log of zero (which should be prevented anyway by dithering).
-  mel_energies_.ApplyFloor(std::numeric_limits<BaseFloat>::epsilon());
+  mel_energies_.ApplyFloor(std::numeric_limits<float>::epsilon());
   mel_energies_.ApplyLog();  // take the log.
 
   feature->SetZero();  // in case there were NaNs.

src/feat/feature-mfcc.h

@@ -40,7 +40,8 @@ struct MfccOptions {
   MelBanksOptions mel_opts;
   int32 num_ceps;  // e.g. 13: num cepstral coeffs, counting zero.
   bool use_energy;  // use energy; else C0
-  BaseFloat energy_floor;
+  BaseFloat energy_floor;  // 0 by default; set to a value like 1.0 or 0.1 if
+                           // you disable dithering.
   bool raw_energy;  // If true, compute energy before preemphasis and windowing
   BaseFloat cepstral_lifter;  // Scaling factor on cepstra for HTK compatibility.
                               // if 0.0, no liftering is done.
@@ -53,7 +54,7 @@ struct MfccOptions {
                   // but for 8khz-sampled data, 15 may be better.
                   num_ceps(13),
                   use_energy(true),
-                  energy_floor(0.0),  // not in log scale: a small value e.g. 1.0e-10
+                  energy_floor(0.0),
                   raw_energy(true),
                   cepstral_lifter(22.0),
                   htk_compat(false) {}
@@ -66,7 +67,9 @@ struct MfccOptions {
     opts->Register("use-energy", &use_energy,
                    "Use energy (not C0) in MFCC computation");
     opts->Register("energy-floor", &energy_floor,
-                   "Floor on energy (absolute, not relative) in MFCC computation");
+                   "Floor on energy (absolute, not relative) in MFCC computation. "
+                   "Only makes a difference if --use-energy=true; only necessary if "
+                   "--dither=0.0.  Suggested values: 0.1 or 1.0");
     opts->Register("raw-energy", &raw_energy,
                    "If true, compute energy before preemphasis and windowing");
     opts->Register("cepstral-lifter", &cepstral_lifter,

src/feat/feature-plp.cc

@@ -125,7 +125,7 @@ void PlpComputer::Compute(BaseFloat signal_log_energy,
 
   if (opts_.use_energy && !opts_.raw_energy)
     signal_log_energy = Log(std::max(VecVec(*signal_frame, *signal_frame),
-                                     std::numeric_limits<BaseFloat>::min()));
+                                     std::numeric_limits<float>::min()));
 
   if (srfft_ != NULL)  // Compute FFT using split-radix algorithm.
     srfft_->Compute(signal_frame->Data(), true);
@@ -160,7 +160,7 @@ void PlpComputer::Compute(BaseFloat signal_log_energy,
   BaseFloat residual_log_energy = ComputeLpc(autocorr_coeffs_, &lpc_coeffs_);
 
   residual_log_energy = std::max(residual_log_energy,
-                                 std::numeric_limits<BaseFloat>::min());
+                                 std::numeric_limits<float>::min());
 
   Lpc2Cepstrum(opts_.lpc_order, lpc_coeffs_.Data(), raw_cepstrum_.Data());
   feature->Range(1, opts_.num_ceps - 1).CopyFromVec(

src/feat/feature-plp.h

@@ -61,7 +61,7 @@ struct PlpOptions {
                  lpc_order(12),
                  num_ceps(13),
                  use_energy(true),
-                 energy_floor(0.0),  // not in log scale: a small value e.g. 1.0e-10
+                 energy_floor(0.0),
                  raw_energy(true),
                  compress_factor(0.33333),
                  cepstral_lifter(22),
@@ -78,7 +78,9 @@ struct PlpOptions {
     opts->Register("use-energy", &use_energy,
                    "Use energy (not C0) for zeroth PLP feature");
     opts->Register("energy-floor", &energy_floor,
-                   "Floor on energy (absolute, not relative) in PLP computation");
+                   "Floor on energy (absolute, not relative) in MFCC computation. "
+                   "Only makes a difference if --use-energy=true; only necessary if "
+                   "--dither=0.0.  Suggested values: 0.1 or 1.0");
     opts->Register("raw-energy", &raw_energy,
                    "If true, compute energy before preemphasis and windowing");
     opts->Register("compress-factor", &compress_factor,

src/feat/feature-spectrogram.cc

@@ -55,7 +55,7 @@ void SpectrogramComputer::Compute(BaseFloat signal_log_energy,
   // Compute energy after window function (not the raw one)
   if (!opts_.raw_energy)
     signal_log_energy = Log(std::max(VecVec(*signal_frame, *signal_frame),
-                                     std::numeric_limits<BaseFloat>::epsilon()));
+                                     std::numeric_limits<float>::epsilon()));
 
   if (srfft_ != NULL)  // Compute FFT using split-radix algorithm.
     srfft_->Compute(signal_frame->Data(), true);
@@ -67,7 +67,7 @@ void SpectrogramComputer::Compute(BaseFloat signal_log_energy,
   SubVector<BaseFloat> power_spectrum(*signal_frame,
                                       0, signal_frame->Dim() / 2 + 1);
 
-  power_spectrum.ApplyFloor(std::numeric_limits<BaseFloat>::epsilon());
+  power_spectrum.ApplyFloor(std::numeric_limits<float>::epsilon());
   power_spectrum.ApplyLog();
 
   feature->CopyFromVec(power_spectrum);

src/feat/feature-spectrogram.h

@@ -41,13 +41,17 @@ struct SpectrogramOptions {
   bool raw_energy;  // If true, compute energy before preemphasis and windowing
 
   SpectrogramOptions() :
-    energy_floor(0.0),  // not in log scale: a small value e.g. 1.0e-10
+    energy_floor(0.0),
     raw_energy(true) {}
 
   void Register(OptionsItf *opts) {
     frame_opts.Register(opts);
     opts->Register("energy-floor", &energy_floor,
-                   "Floor on energy (absolute, not relative) in Spectrogram computation");
+                   "Floor on energy (absolute, not relative) in Spectrogram "
+                   "computation.  Caution: this floor is applied to the zeroth "
+                   "component, representing the total signal energy.  The "
+                   "floor on the individual spectrogram elements is fixed at "
+                   "std::numeric_limits<float>::epsilon().");
     opts->Register("raw-energy", &raw_energy,
                    "If true, compute energy before preemphasis and windowing");
   }

src/feat/feature-window.cc

@@ -145,7 +145,7 @@ void ProcessWindow(const FrameExtractionOptions &opts,
 
   if (log_energy_pre_window != NULL) {
     BaseFloat energy = std::max(VecVec(*window, *window),
-                                std::numeric_limits<BaseFloat>::epsilon());
+                                std::numeric_limits<float>::epsilon());
     *log_energy_pre_window = Log(energy);
   }
 

src/feat/feature-window.h

@@ -71,7 +71,9 @@ struct FrameExtractionOptions {
                    "Coefficient for use in signal preemphasis");
     opts->Register("remove-dc-offset", &remove_dc_offset,
                    "Subtract mean from waveform on each frame");
-    opts->Register("dither", &dither, "Dithering constant (0.0 means no dither)");
+    opts->Register("dither", &dither, "Dithering constant (0.0 means no dither). "
+                   "If you turn this off, you should set the --energy-floor "
+                   "option, e.g. to 1.0 or 0.1");
     opts->Register("window-type", &window_type, "Type of window "
                    "(\"hamming\"|\"hanning\"|\"povey\"|\"rectangular\""
                    "|\"blackmann\")");