Postfilter functions

Postfilter functions (gain control, wiener denoise filter, DC filter, kalman smoothening, plus surrounding code to wrap it). More...

Functions
static void	adaptive_gain_control (float *out, const float *in, const float *speech_synth, int size, float alpha, float *gain_mem)
	Adaptive gain control (as used in postfilter).
static int	kalman_smoothen (WMAVoiceContext *s, int pitch, const float *in, float *out, int size)
	Kalman smoothing function.
static float	tilt_factor (const float *lpcs, int n_lpcs)
	Get the tilt factor of a formant filter from its transfer function.
static void	calc_input_response (WMAVoiceContext *s, float *lpcs, int fcb_type, float *coeffs, int remainder)
	Derive denoise filter coefficients (in real domain) from the LPCs.
static void	wiener_denoise (WMAVoiceContext *s, int fcb_type, float *synth_pf, int size, const float *lpcs)
	This function applies a Wiener filter on the (noisy) speech signal as a means to denoise it.
static void	postfilter (WMAVoiceContext *s, const float *synth, float *samples, int size, const float *lpcs, float *zero_exc_pf, int fcb_type, int pitch)
	Averaging projection filter, the postfilter used in WMAVoice.

---

Detailed Description

Postfilter functions (gain control, wiener denoise filter, DC filter, kalman smoothening, plus surrounding code to wrap it).

---

Function Documentation

static void adaptive_gain_control	(	float *	out,
		const float *	in,
		const float *	speech_synth,
		int	size,
		float	alpha,
		float *	gain_mem
	)			[static]

Adaptive gain control (as used in postfilter).

Identical to ff_adaptive_gain_control() in acelp_vectors.c, except that the energy here is calculated using sum(abs(...)), whereas the other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).

Parameters:

	out	output buffer for filtered samples
	in	input buffer containing the samples as they are after the postfilter steps so far
	speech_synth	input buffer containing speech synth before postfilter
	size	input buffer size
	alpha	exponential filter factor
	gain_mem	pointer to filter memory (single float)

Definition at line 454 of file wmavoice.c.

Referenced by postfilter().

static void calc_input_response	(	WMAVoiceContext *	s,
		float *	lpcs,
		int	fcb_type,
		float *	coeffs,
		int	remainder
	)			[static]

Derive denoise filter coefficients (in real domain) from the LPCs.

Definition at line 553 of file wmavoice.c.

Referenced by wiener_denoise().

static int kalman_smoothen	(	WMAVoiceContext *	s,
		int	pitch,
		const float *	in,
		float *	out,
		int	size
	)			[static]

Kalman smoothing function.

This function looks back pitch +/- 3 samples back into history to find the best fitting curve (that one giving the optimal gain of the two signals, i.e. the highest dot product between the two), and then uses that signal history to smoothen the output of the speech synthesis filter.

Parameters:

	s	WMA Voice decoding context
	pitch	pitch of the speech signal
	in	input speech signal
	out	output pointer for smoothened signal
	size	input/output buffer size

Returns:

-1 if no smoothening took place, e.g. because no optimal fit could be found, or 0 on success.

Definition at line 494 of file wmavoice.c.

Referenced by postfilter().

static void postfilter	(	WMAVoiceContext *	s,
		const float *	synth,
		float *	samples,
		int	size,
		const float *	lpcs,
		float *	zero_exc_pf,
		int	fcb_type,
		int	pitch
	)			[static]

Averaging projection filter, the postfilter used in WMAVoice.

This uses the following steps:

• A zero-synthesis filter (generate excitation from synth signal)
• Kalman smoothing on excitation, based on pitch
• Re-synthesized smoothened output
• Iterative Wiener denoise filter
• Adaptive gain filter
• DC filter

Parameters:

	s	WMAVoice decoding context
	synth	Speech synthesis output (before postfilter)
	samples	Output buffer for filtered samples
	size	Buffer size of synth & samples
	lpcs	Generated LPCs used for speech synthesis
	fcb_type	Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
	pitch	Pitch of the input signal

Definition at line 750 of file wmavoice.c.

static float tilt_factor	(	const float *	lpcs,
		int	n_lpcs
	)			[static]

Get the tilt factor of a formant filter from its transfer function.

See also:

tilt_factor() in amrnbdec.c, which does essentially the same, but somehow (??) it does a speech synthesis filter in the middle, which is missing here

Parameters:

	lpcs	LPC coefficients
	n_lpcs	Size of LPC buffer

Returns:

the tilt factor

Definition at line 540 of file wmavoice.c.

static void wiener_denoise	(	WMAVoiceContext *	s,
		int	fcb_type,
		float *	synth_pf,
		int	size,
		const float *	lpcs
	)			[static]

This function applies a Wiener filter on the (noisy) speech signal as a means to denoise it.

• take RDFT of LPCs to get the power spectrum of the noise + speech;
• using this power spectrum, calculate (for each frequency) the Wiener filter gain, which depends on the frequency power and desired level of noise subtraction (when set too high, this leads to artifacts) We can do this symmetrically over the X-axis (so 0-4kHz is the inverse of 4-8kHz);
• by doing a phase shift, calculate the Hilbert transform of this array of per-frequency filter-gains to get the filtering coefficients;
• smoothen/normalize/de-tilt these filter coefficients as desired;
• take RDFT of noisy sound, apply the coefficients and take its IRDFT to get the denoised speech signal;
• the leftover (i.e. output of the IRDFT on denoised speech data beyond the frame boundary) are saved and applied to subsequent frames by an overlap-add method (otherwise you get clicking-artifacts).

Parameters:

	s	WMA Voice decoding context
	s	fcb_type Frame (codebook) type
	synth_pf	input: the noisy speech signal, output: denoised speech data; should be 16-byte aligned (for ASM purposes)
	size	size of the speech data
	lpcs	LPCs used to synthesize this frame's speech data

Definition at line 669 of file wmavoice.c.

Referenced by postfilter().