1. ON THE ARCHITECTURE OF THE CDMA2000® VARIABLE-RATE MULTIMODE WIDEBAND (VMR-WB) SPEECH CODING STANDARD
Milan Jelinek†, Redwan Salami‡, Sassan Ahmadi*, Bruno Bessette†, Philippe Gournay†‡ and Claude Laflamme†
†University of Sherbrooke, Canada - ‡VoiceAge Corp., Canada - *Nokia inc., USA
VMR-WB
Variable-Rate Multi-Mode Wideband Speech Codec
New 3GPP2 WB Speech Coding Standard for 3G applications
Main Features:
Near Face-to-Face Communication Speech Quality
Source and Channel Controlled Operation (4 Modes)
3GPP/ITU AMR-WB Directly Interoperable in Mode 3
Average Bit Rates (ABR):
Compliant with CDMA2000 Rate Set 2
(FR), 6.2 (HR) , 2.7 (QR) or 1.0 (ER) kbit/s frames
WB ( HZ) and NB ( Hz) Input/Output
20 ms Frames
Noise Reduction with Adjustable Maximum Reduction
Encoder Flow Chart
VMR-WB Coding Techniques
Source-Controlled Operation
Hierarchical Signal Classification
Operating on Frame-level
1. Voice Activity Detection (VAD)
2. Unvoiced Frame Decision
Spectral Analysis
LP Analysis
Pitch Tracking
Noise Reduction
Noise Estimation
Voice Activity?
Voice Activity Decision:
Parameters
Input
De-noised
lower for noisy speech
higher for clean speech
Based on the following parameters:
Coding Type
Bitrate
kbit/s
Description
Inactive Speech Coding
CNG ER
1.0
-Noise excited LP filter
-Smoothed over time
CNG QR
2.7
-As previous, but interoperable with AMR-WB CNG
Unvoiced Coding
Unvoiced HR
6.2
-13 bit Gaussian codebook (4x/frame)
Unvoiced QR
-As previous, but randomly chosen vectors
Voiced Coding
Voiced HR
-Frame level signal modification
-12 bit ACELP codebook (4x/frame)
Generic Coding
Interoperable FR
13.3
-Similar to kbit/s
Generic FR
-As previous + FER protection
Interoperable HR
-As Interoperable FR, but with random algebraic codebook indices
Signaling HR
Generic HR
-Pitch coded 2x/frame
Normalized Correlation
T – open-loop pitch period estimate
xi – perceptually weighted input signal
Begin
1. Voice Activity?
2. Unvoiced Frame?
3. Voiced Frame?
4. Low Energy?
CNG Encoding or DTX
Unvoiced Speech Optimized Encoding
Voiced Speech Optimized Encoding
Generic HR Encoding
Generic FR Encoding
Yes No
Spectral Tilt
Eh – average energy of last 2 critical bands.
El – average energy of pitch-synchronous bins in the first 10 critical bands
Active speech
kbit/s
40% Speech Activity
Mode 3
13.3
6.1
Mode 0
12.8
5.7
Mode 1
10.5
4.8
Mode 2
8.1
3.8
Frame Energy Variation
Noise Estimation Update Decision:
Based on parameters with low sensitivity to noise level:
Pitch period varying
AND normalized correlation at pitch period low
AND low estimated order of AR model
AND signal energy stationary
INDEPENDENT of VAD decision!
- Robust to noise level variations
- Conservative approach: the noise estimation is updated only if quite sure the frame is inactive
E32(j) – energy maximum in a bloc of 32-samples
Relative Frame Energy - Erel
Decision:
3. Voiced Frame Decision / Signal Modification
4. Low Energy Decision
Channel-Controlled Operation
4 Operational Modes Controlled by Channel Conditions
Transparent Memory-less Mode Switching
Per-Frame Bit Rate Control Capability
  Coding Types Relative Usage in Active Speech:
Mode Switching Performance:
Comparing MOS scores of modes 0, 1, 2 with random mode switching at 0.5, 1 and 5 second intervals (from characterization test)
Enhancements at Decoder
Low Frequency Post-processing:
Enhancement of the periodicity in low frequency region:
Performance (MOS scores from selection test)
CDMA Specific Modes (Modes 0, 1, 2), WB Input
Performance (MOS scores from characterization test)
Voiced Decision is an Inherent Part of Original Signal Modification Algorithm
Frame is coded as voiced if all constraints of the modification are satisfied
Signal modification is done pitch-synchronously
Pitch period evolution is piecewise linear (constant at frame end) to avoid pitch period oscillations
Modified input is synchronous with original input at frame end
Modification is transparent at least up to 30% of active speech frames (in the example bellow, no coding is used and 30 % of active clean speech frames are modified)
NB Input Test
Modes 0, 1, 2, 3,
Clean speech, nominal level
Test on Interworking with kbit/s
-WB input, clean speech conditions
Purpose:
To avoid encoding unclassified frames with low perceptual importance at Full Rate
Condition:
2000 Hz
Ref 0 – 14.25
Ref 1 – 12.65
Ref 2 – 8.85
Test 0 – VMR-WB Mode 0
Test 1 – VMR-WB Mode 1
Test 2 – VMR-WB Mode 2
Coding Type
Mode 0
Mode 1
Mode 2
Mode 3
Generic FR
93.4 %
60.4 %
34.1 % -
Interoperable FR
100.0 %
Generic HR
7.1 %
13.1 %
Voiced HR
13.0 %
33.2 %
Unvoiced HR
6.6 %
19.5 %
5.6 %
Unvoiced QR
14.0 %
Et – sum of critical band energies for current frame, in dB
Ef – long-term mean of Et for active speech
Clean Speech
Conditions:
Example:
Typical example of a low-energy frame encoded with Generic HR in mode 2
Frame Errors Concealment:
Lost Frame Concealment:
Excitation energy and spectral envelope converge to estimated noise.
Excitation periodicity converges to 0.
Convergence rate depends on the signal class of last good frame.
Recovery after erasure:
Careful energy control of synthesized speech.
Artificial onset reconstruction in case of lost voiced onset.
Channel Error
Conditions:
Background Noise
Conditions: