Acoustic model structures for PocketSphinx. More...

#include <assert.h>
#include <string.h>
#include <math.h>
#include <sphinxbase/prim_type.h>
#include <sphinxbase/err.h>
#include <sphinxbase/cmd_ln.h>
#include <sphinxbase/strfuncs.h>
#include <sphinxbase/byteorder.h>
#include <sphinxbase/feat.h>
#include <sphinxbase/bio.h>
#include "cmdln_macro.h"
#include "acmod.h"
#include "s2_semi_mgau.h"
#include "ptm_mgau.h"
#include "ms_mgau.h"

Macros
#define	MPX_BITVEC_SET(a, h, i)

#define	NONMPX_BITVEC_SET(a, h, i)

Functions
int	acmod_fe_mismatch (acmod_t acmod, fe_t fe)

int	acmod_feat_mismatch (acmod_t acmod, feat_t fcb)

acmod_t *	acmod_init (cmd_ln_t config, logmath_t lmath, fe_t fe, feat_t fcb)
	Initialize an acoustic model. More...

void	acmod_free (acmod_t *acmod)
	Finalize an acoustic model.

ps_mllr_t *	acmod_update_mllr (acmod_t acmod, ps_mllr_t mllr)
	Adapt acoustic model using a linear transform. More...

int	acmod_write_senfh_header (acmod_t acmod, FILE logfh)
	Write senone dump file header.

int	acmod_set_senfh (acmod_t acmod, FILE logfh)
	Start logging senone scores to a filehandle. More...

int	acmod_set_mfcfh (acmod_t acmod, FILE logfh)
	Start logging MFCCs to a filehandle. More...

int	acmod_set_rawfh (acmod_t acmod, FILE logfh)
	Start logging raw audio to a filehandle. More...

void	acmod_grow_feat_buf (acmod_t *acmod, int nfr)

int	acmod_set_grow (acmod_t *acmod, int grow_feat)
	Set memory allocation policy for utterance processing. More...

int	acmod_start_utt (acmod_t *acmod)
	Mark the start of an utterance.

int	acmod_end_utt (acmod_t *acmod)
	Mark the end of an utterance.

int	acmod_process_raw (acmod_t acmod, int16 const inout_raw, size_t inout_n_samps, int full_utt)
	TODO: Set queue length for utterance processing. More...

int	acmod_process_cep (acmod_t acmod, mfcc_t *inout_cep, int inout_n_frames, int full_utt)
	Feed acoustic feature data into the acoustic model for scoring. More...

int	acmod_process_feat (acmod_t acmod, mfcc_t *feat)
	Feed dynamic feature data into the acoustic model for scoring. More...

int	acmod_set_insenfh (acmod_t acmod, FILE senfh)
	Set up a senone score dump file for input. More...

int	acmod_rewind (acmod_t *acmod)
	Rewind the current utterance, allowing it to be rescored. More...

int	acmod_advance (acmod_t *acmod)
	Advance the frame index. More...

int	acmod_write_scores (acmod_t acmod, int n_active, uint8 const active, int16 const senscr, FILE senfh)
	Write a frame of senone scores to a dump file.

int	acmod_read_scores (acmod_t *acmod)
	Read one frame of scores from senone score dump file. More...

mfcc_t **	acmod_get_frame (acmod_t acmod, int inout_frame_idx)
	Get a frame of dynamic feature data. More...

int16 const *	acmod_score (acmod_t acmod, int inout_frame_idx)
	Score one frame of data. More...

int	acmod_best_score (acmod_t acmod, int out_best_senid)
	Get best score and senone index for current frame.

void	acmod_clear_active (acmod_t *acmod)
	Clear set of active senones.

void	acmod_activate_hmm (acmod_t acmod, hmm_t hmm)
	Activate senones associated with an HMM.

int32	acmod_flags2list (acmod_t *acmod)
	Build active list from.

int32	acmod_stream_offset (acmod_t *acmod)
	Get the offset of the utterance start of the current stream, helpful for stream-wide timing.

void	acmod_start_stream (acmod_t *acmod)
	Reset the current stream.

void	acmod_set_rawdata_size (acmod_t *acmod, int32 size)
	Sets the limit of the raw audio data to store.

void	acmod_get_rawdata (acmod_t acmod, int16 buffer, int32 size)
	Retrieves the raw data collected during utterance decoding.

Detailed Description

Acoustic model structures for PocketSphinx.

Author: David Huggins-Daines dhugg.nosp@m.ins@.nosp@m.cs.cm.nosp@m.u.ed.nosp@m.u

Definition in file acmod.c.

Macro Definition Documentation

#define MPX_BITVEC_SET	(	a,
		h,
		i
	)

Value:

if (hmm_mpx_ssid(h,i) != BAD_SSID) \

bitvec_set((a)->senone_active_vec, hmm_mpx_senid(h,i))

BAD_SSID

#define BAD_SSID

Invalid senone sequence ID (limited to 16 bits for PocketSphinx).

Definition: bin_mdef.h:94

Definition at line 1205 of file acmod.c.

#define NONMPX_BITVEC_SET	(	a,
		h,
		i
	)

Value:

bitvec_set((a)->senone_active_vec, \

hmm_nonmpx_senid(h,i))

Definition at line 1208 of file acmod.c.

Function Documentation

int acmod_advance ( acmod_t * acmod )

Advance the frame index.

This function moves to the next frame of input data. Subsequent calls to acmod_score() will return scores for that frame, until the next call to acmod_advance().

Returns: New frame index.

Definition at line 899 of file acmod.c.

References acmod_s::feat_outidx, ps_mgau_s::frame_idx, acmod_s::mgau, acmod_s::n_feat_alloc, acmod_s::n_feat_frame, and acmod_s::output_frame.

mfcc_t** acmod_get_frame	(	acmod_t *	acmod,
		int *	inout_frame_idx
	)

Get a frame of dynamic feature data.

Parameters

inout_frame_idx Input: frame index to get, or NULL to obtain features for the most recent frame. Output: frame index corresponding to this set of features.

Returns: Feature array, or NULL if requested frame is not available.

Definition at line 1088 of file acmod.c.

References acmod_s::feat_buf.

acmod_t* acmod_init	(	cmd_ln_t *	config,
		logmath_t *	lmath,
		fe_t *	fe,
		feat_t *	fcb
	)

Initialize an acoustic model.

Parameters

config	a command-line object containing parameters. This pointer is not retained by this object.
lmath	global log-math parameters.
fe	a previously-initialized acoustic feature module to use, or NULL to create one automatically. If this is supplied and its parameters do not match those in the acoustic model, this function will fail. This pointer is not retained.
fe	a previously-initialized dynamic feature module to use, or NULL to create one automatically. If this is supplied and its parameters do not match those in the acoustic model, this function will fail. This pointer is not retained.

Returns: a newly initialized acmod_t, or NULL on failure.

Definition at line 228 of file acmod.c.

References acmod_free(), ACMOD_IDLE, acmod_s::compallsen, acmod_s::config, acmod_s::fcb, acmod_s::fe, acmod_s::feat_buf, acmod_s::framepos, acmod_s::lmath, acmod_s::log_zero, acmod_s::mdef, acmod_s::mfc_buf, acmod_s::n_feat_alloc, acmod_s::n_mfc_alloc, acmod_s::senone_active, acmod_s::senone_active_vec, acmod_s::senone_scores, acmod_s::state, and acmod_s::utt_start_frame.

Referenced by ps_reinit().

int acmod_process_cep	(	acmod_t *	acmod,
		mfcc_t ***	inout_cep,
		int *	inout_n_frames,
		int	full_utt
	)

Feed acoustic feature data into the acoustic model for scoring.

Parameters

inout_cep	In: Pointer to buffer of features Out: Pointer to next frame to be read
inout_n_frames	In: Number of frames available Out: Number of frames remaining
full_utt	If non-zero, this block represents a full utterance and should be processed as such.

Returns: Number of frames of data processed.

Definition at line 699 of file acmod.c.

References ACMOD_ENDED, ACMOD_PROCESSING, ACMOD_STARTED, acmod_s::fcb, acmod_s::feat_buf, acmod_s::feat_outidx, acmod_s::grow_feat, acmod_s::mfcfh, acmod_s::n_feat_alloc, acmod_s::n_feat_frame, and acmod_s::state.

int acmod_process_feat	(	acmod_t *	acmod,
		mfcc_t **	feat
	)

Feed dynamic feature data into the acoustic model for scoring.

Unlike acmod_process_raw() and acmod_process_cep(), this function accepts a single frame at a time. This is because there is no need to do buffering when using dynamic features as input. However, if the dynamic feature buffer is full, this function will fail, so you should either always check the return value, or always pair a call to it with a call to acmod_score().

Parameters

feat	Pointer to one frame of dynamic features.

Returns: Number of frames processed (either 0 or 1).

Definition at line 797 of file acmod.c.

References acmod_s::fcb, acmod_s::feat_buf, acmod_s::feat_outidx, acmod_s::grow_feat, acmod_s::n_feat_alloc, and acmod_s::n_feat_frame.

int acmod_process_raw	(	acmod_t *	acmod,
		int16 const **	inout_raw,
		size_t *	inout_n_samps,
		int	full_utt
	)

TODO: Set queue length for utterance processing.

This function allows multiple concurrent passes of search to operate on different parts of the utterance. Feed raw audio data to the acoustic model for scoring.

Parameters

inout_raw	In: Pointer to buffer of raw samples Out: Pointer to next sample to be read
inout_n_samps	In: Number of samples available Out: Number of samples remaining
full_utt	If non-zero, this block represents a full utterance and should be processed as such.

Returns: Number of frames of data processed.

Definition at line 607 of file acmod.c.

References acmod_s::fe, acmod_s::mfc_buf, acmod_s::mfc_outidx, acmod_s::n_mfc_alloc, acmod_s::n_mfc_frame, acmod_s::rawfh, and acmod_s::utt_start_frame.

Referenced by ps_process_raw().

int acmod_read_scores ( acmod_t * acmod )

Read one frame of scores from senone score dump file.

Returns: Number of frames read or <0 on error.

Definition at line 1012 of file acmod.c.

References acmod_s::feat_outidx, acmod_s::framepos, acmod_s::grow_feat, acmod_s::insenfh, acmod_s::n_feat_alloc, acmod_s::n_feat_frame, acmod_s::n_senone_active, acmod_s::output_frame, and acmod_s::senscr_frame.

Referenced by ps_decode_senscr().

int acmod_rewind ( acmod_t * acmod )

Rewind the current utterance, allowing it to be rescored.

After calling this function, the internal frame index is reset, and acmod_score() will return scores starting at the first frame of the current utterance. Currently, acmod_set_grow() must have been called to enable growing the feature buffer in order for this to work. In the future, senone scores may be cached instead.

Returns: 0 for success, <0 for failure (if the utterance can't be rewound due to no feature or score data available)

Definition at line 877 of file acmod.c.

References acmod_s::feat_outidx, ps_mgau_s::frame_idx, acmod_s::mgau, acmod_s::n_feat_alloc, acmod_s::n_feat_frame, acmod_s::output_frame, and acmod_s::senscr_frame.

int16 const* acmod_score	(	acmod_t *	acmod,
		int *	inout_frame_idx
	)

Score one frame of data.

Parameters

inout_frame_idx Input: frame index to score, or NULL to obtain scores for the most recent frame. Output: frame index corresponding to this set of scores.

Returns: Array of senone scores for this frame, or NULL if no frame is available for scoring (such as if a frame index is requested that is not yet or no longer available). The data pointed to persists only until the next call to acmod_score() or acmod_advance().

Definition at line 1106 of file acmod.c.

References acmod_flags2list(), acmod_write_scores(), acmod_s::compallsen, acmod_s::feat_buf, acmod_s::framepos, acmod_s::insenfh, acmod_s::mgau, acmod_s::n_senone_active, acmod_s::senfh, acmod_s::senone_active, acmod_s::senone_scores, and acmod_s::senscr_frame.

Referenced by ngram_fwdflat_search(), and ngram_fwdtree_search().

int acmod_set_grow	(	acmod_t *	acmod,
		int	grow_feat
	)

Set memory allocation policy for utterance processing.

Parameters

grow_feat If non-zero, the internal dynamic feature buffer will expand as necessary to encompass any amount of data fed to the model.

Returns: previous allocation policy.

Definition at line 410 of file acmod.c.

References acmod_s::grow_feat, and acmod_s::n_feat_alloc.

Referenced by ngram_search_init(), and ps_process_raw().

int acmod_set_insenfh	(	acmod_t *	acmod,
		FILE *	insenfh
	)

Set up a senone score dump file for input.

Parameters

insenfh File handle of dump file

Returns: 0 for success, <0 for failure

Definition at line 864 of file acmod.c.

References acmod_s::compallsen, acmod_s::config, acmod_s::insenfh, and acmod_s::n_feat_frame.

Referenced by ps_decode_senscr().

int acmod_set_mfcfh	(	acmod_t *	acmod,
		FILE *	logfh
	)

Start logging MFCCs to a filehandle.

Parameters

acmod	Acoustic model object.
logfh	Filehandle to log to.

Returns: 0 for success, <0 on error.

Definition at line 375 of file acmod.c.

References acmod_s::mfcfh.

Referenced by ps_start_utt().

int acmod_set_rawfh	(	acmod_t *	acmod,
		FILE *	logfh
	)

Start logging raw audio to a filehandle.

Parameters

acmod	Acoustic model object.
logfh	Filehandle to log to.

Returns: 0 for success, <0 on error.

Definition at line 387 of file acmod.c.

References acmod_s::rawfh.

Referenced by ps_start_utt().

int acmod_set_senfh	(	acmod_t *	acmod,
		FILE *	senfh
	)

Start logging senone scores to a filehandle.

Parameters

acmod	Acoustic model object.
logfh	Filehandle to log to.

Returns: 0 for success, <0 on error.

Definition at line 364 of file acmod.c.

References acmod_write_senfh_header(), and acmod_s::senfh.

Referenced by ps_start_utt().

ps_mllr_t* acmod_update_mllr	(	acmod_t *	acmod,
		ps_mllr_t *	mllr
	)

Adapt acoustic model using a linear transform.

Parameters

mllr	The new transform to use, or NULL to update the existing transform. The decoder retains ownership of this pointer, so you should not attempt to free it manually. Use ps_mllr_retain() if you wish to reuse it elsewhere.

Returns: The updated transform object for this decoder, or NULL on failure.

Definition at line 339 of file acmod.c.

References acmod_s::mgau, acmod_s::mllr, and ps_mllr_free().

Referenced by ps_update_mllr().

Macros

Functions

Detailed Description

Macro Definition Documentation

Function Documentation