1. Philippe Gournay, Bruno Bessette, Roch Lefebvre
Universal Speech and Audio Codec
Linear Prediction Domain processing
Philippe Gournay, Bruno Bessette, Roch Lefebvre
Université de Sherbrooke
Département de Génie Electrique et Informatique
Sherbrooke, Québec, Canada

2. Outline The 3GPP AMR-WB+ Standard Changes brought to LPD processing
Source of inspiration for LPD processing in USAC
Changes brought to LPD processing
Forward Aliasing Cancellation
Frequency-Domain Noise Shaping
Other changes
Conclusion
More efficient LPD processing
Better unification of LPD and non-LPD FD coders

3. Context The 3GPP AMR-WB+ Standard Hybrid codec
Time (ACELP) and Frequency (TCX) Domain
Very efficient on speech and speech-over-music contents

4. The AMR-WB+ Encoder ACELP PACKETIZATION 1 frame Bitstream Audio TCX
Mode Selection
1, 2 or 4 frames
Mode Index, ISF

5. AMR-WB+ Frame Structure
ACELP
Short TCX
Medium TCX
Long TCX
One super-frame = 1024 samples
(a)
(b)
(c)
Three out of the 26 possible ACELP/TCX coding configurations

6. Transitions from ACELP to TCX
Zero-input response (ZIR) of LPC weighting filter provides pseudo-windowing
Decoded TCX window
ACELP
Frame
1/8 overlap

7. Transitions from TCX to ACELP
Redundant windowed TCX samples are discarded
Decoded TCX window
Frame
1/8 overlap
ACELP

8. Limitations of the AMR-WB+ model
Non-critically sampled transforms
FFT vs. MDCT
Inefficiencies at transitions between modes
Sub-optimal windowing (from ACELP to TCX)
Discarded samples (from TCX to ACELP)
Transform windows not aligned with ACELP grid
LPC analysis window also shifted to the right
Even worse when switching with AAC
Time-Domain Aliasing Cancellation (TDAC)
Transitions between LPD and non-LPD processing

9. Changes brought to the LPD processing
Replaced FFTs by MDCTs
Introduced Frequency Domain Noise Shaping
Introduced Forward Aliasing Cancellation
Other changes

10. Frequency Domain Noise Shaping
To unify processing of AAC and TCX frames, the MDCT transform in TCX is applied in the original signal domain
Noise shaping for TCX frames is performed in the MDCT domain based on LPC filters mapped to the MDCT domain
FDNS allows a smooth (sample-by-sample) time-domain noise envelope by applying a 1st-order filtering to the MDCT coefficients (similar in principle to TNS)

11. Effect of FDNS on the spectral shape and the time-domain envelope of the noise
time axis (n) A B C
Noise gains
g1[m] calculated at time position A
Interpolated gains seen
in the time domain, for each
of the M bands
Noise gains g2[m] calculated at time position B
Frequency axis
(k or m)

12. Frequency-Domain Noise Shaping
FDNS allows a smooth (sample-by-sample) time-domain noise envelope by applying a 1st-order filtering to the MDCT coefficients (similar in principle to TNS)

13. Forward Aliasing Cancellation
Introduced to compensate windowing and time-domain aliasing in MDCT-coded frames when switching to and from ACELP frames
Windowing effect and
Time Domain Aliasing
ACELP synthesis
TCX frame output
Next ACELP frame - +

14. Forward Aliasing Cancellation
FAC is applied in the original signal domain
FAC is quantized in the LPC weighted domain so that quantization noises of FAC and decoded MDCT are of the same nature
For transition from ACELP to TCX, the ACELP synthesis can be taken into account; this reduces the bitrate needed to encode FAC

15. Computation of FAC targets for transitions from and to ACELP (encoder)
Signal in the original domain + -
TCX frame output
ACELP synthesis
ACELP contribution
TCX frame error
(including ACELP contribution)
ACELP error
LPC1
LPC2
Windowed and folded
ACELP synth
Windowed ACELP ZIR
Line 1
Line 2
Line 3
Line 4
Next ACELP frame
FAC target

16. Quantization of FAC targets
W1(z)
LPC1
FAC target
DCT-IV Q
DCT-IV-1
FAC synthesis
1/W1(z)
1/W1(z) ZIR
Transmit to decoder
Filter memory
(ACELP error)
Zero memory
Transition from ACELP to TCX
LPC2
FAC target
FAC synthesis
W2(z)
DCT-IV Q
DCT-IV-1
1/W2(z)
Filter memory
(TCX frame error)
Zero memory
Transmit to decoder
Transition from TC to ACELP

17. Other changes brought to the LPD processing
Critical sampling
MDCT vs. FFT
FAC+FDNS
Scalar quantizer + adaptive arithmetic coder for TCX (AMR-WB+ uses AVQ)
Variable bit rate
LPC quantizer
Bit reservoir adaptation

18. Conclusion USAC makes use of LPD and non-LPD processing
LPD mode inspired by AMR-WB+
Non-LPD mode derived from AAC
Substantial changes were brought to the LPD processing, and new tools were introduced to make it more efficient
Frequency Domain Noise Shaping (FDNS)
Forward Aliasing Cancellation (FAC)
USAC is a real unification of two coding models

19. Thank you for your attention!