This configuration class controls when to re-estimate the basis-fMLLR during online decoding. More...

#include <online-gmm-decoding.h>

Collaboration diagram for OnlineGmmDecodingAdaptationPolicyConfig:

Public Member Functions
	OnlineGmmDecodingAdaptationPolicyConfig ()

void	Register (OptionsItf *opts)

void	Check () const
	Check that configuration values make sense. More...

bool	DoAdapt (BaseFloat chunk_begin_secs, BaseFloat chunk_end_secs, bool is_first_utterance) const
	This function returns true if we are scheduled to re-estimate fMLLR somewhere in the interval [ chunk_begin_secs, chunk_end_secs ). More...

Public Attributes
BaseFloat	adaptation_first_utt_delay

BaseFloat	adaptation_first_utt_ratio

BaseFloat	adaptation_delay

BaseFloat	adaptation_ratio

Detailed Description

This configuration class controls when to re-estimate the basis-fMLLR during online decoding.

The basic model is to re-estimate it on a certain time t (e.g. after 1 second) and then at a set of times forming a geometric series, e.g. 1.5, 1.5^2, etc. We specify different configurations for the first utterance of a speaker (which requires more frequent adaptation), and for subsequent utterances. We also re-estimate fMLLR at the end of every utterance, but this is done directly from the calling code, not by the class SingleUtteranceGmmDecoder.

Definition at line 56 of file online-gmm-decoding.h.

Constructor & Destructor Documentation

◆ OnlineGmmDecodingAdaptationPolicyConfig()

OnlineGmmDecodingAdaptationPolicyConfig ( )

inline

Definition at line 61 of file online-gmm-decoding.h.

                                            :
       adaptation_first_utt_delay(2.0),
       adaptation_first_utt_ratio(1.5),
       adaptation_delay(5.0),
       adaptation_ratio(2.0) { }

Member Function Documentation

◆ Check()

void Check ( ) const

Check that configuration values make sense.

Definition at line 416 of file online-gmm-decoding.cc.

References KALDI_ASSERT.

Referenced by OnlineGmmDecodingAdaptationPolicyConfig::Register().

                                                           {
   KALDI_ASSERT(adaptation_first_utt_delay > 0.0 &&
                adaptation_first_utt_ratio > 1.0);
   KALDI_ASSERT(adaptation_delay > 0.0 &&
                adaptation_ratio > 1.0);
 }

◆ DoAdapt()

bool DoAdapt	(	BaseFloat	chunk_begin_secs,
		BaseFloat	chunk_end_secs,
		bool	is_first_utterance
	)		const

This function returns true if we are scheduled to re-estimate fMLLR somewhere in the interval [ chunk_begin_secs, chunk_end_secs ).

Definition at line 423 of file online-gmm-decoding.cc.

Referenced by SingleUtteranceGmmDecoder::AdvanceDecoding(), and OnlineGmmDecodingAdaptationPolicyConfig::Register().

                                    {
   Check();
   if (is_first_utterance) {
     // We aim to return true if a member of the sequence
     // ( adaptation_first_utt_delay * adaptation_first_utt_ratio^n )
     // for  n = 0, 1, 2, ...
     // is in the range [ chunk_begin_secs, chunk_end_secs ).
     BaseFloat delay = adaptation_first_utt_delay;
     while (delay < chunk_begin_secs)
       delay *= adaptation_first_utt_ratio;
     return (delay < chunk_end_secs);
   } else {
     // as above, but remove "first_utt".
     BaseFloat delay = adaptation_delay;
     while (delay < chunk_begin_secs)
       delay *= adaptation_ratio;
     return (delay < chunk_end_secs);
   }
 }

◆ Register()

void Register ( OptionsItf * opts )

inline

Definition at line 67 of file online-gmm-decoding.h.

References OnlineGmmDecodingAdaptationPolicyConfig::Check(), OnlineGmmDecodingAdaptationPolicyConfig::DoAdapt(), and OptionsItf::Register().

Referenced by OnlineGmmDecodingConfig::Register().

                                   {
     opts->Register("adaptation-first-utt-delay", &adaptation_first_utt_delay,
                    "Delay before first basis-fMLLR adaptation for first utterance "
                    "of each speaker");
     opts->Register("adaptation-first-utt-ratio", &adaptation_first_utt_ratio,
                    "Ratio that controls frequency of fMLLR adaptation for first "
                    "utterance of each speaker");
     opts->Register("adaptation-delay", &adaptation_delay,
                    "Delay before first basis-fMLLR adaptation for not-first "
                    "utterances of each speaker");
     opts->Register("adaptation-ratio", &adaptation_ratio,
                    "Ratio that controls frequency of fMLLR adaptation for "
                    "not-first utterances of each speaker");
   }

Member Data Documentation

◆ adaptation_delay

BaseFloat adaptation_delay

Definition at line 59 of file online-gmm-decoding.h.

◆ adaptation_first_utt_delay

BaseFloat adaptation_first_utt_delay

Definition at line 57 of file online-gmm-decoding.h.

◆ adaptation_first_utt_ratio

BaseFloat adaptation_first_utt_ratio

Definition at line 58 of file online-gmm-decoding.h.

◆ adaptation_ratio

BaseFloat adaptation_ratio

Definition at line 60 of file online-gmm-decoding.h.

The documentation for this struct was generated from the following files:

online2/online-gmm-decoding.h
online2/online-gmm-decoding.cc

Public Member Functions

Public Attributes