Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 1

IEEE Speech Coding WorkshopSept 17–20, 2000Lake Lawn Resort

Delavan, WI

Jean-Marc Valin, Roch LefebvreUniversity of Sherbrooke

Bandwidth Extension of Narrowband Speech for Low Bit-Rate Wideband Coding

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 2

Outline

• Problem statement

• Proposed solution

• System performance

• Discussion

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 3

Problem Statement

• Telephone Band: 300 - 3400 Hz

• AM Band: 50 - 7000 Hz

• How to make sound like with 500 bits/sec?

• We need to recover information from both low and high-frequency bands

(G.729)

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 4

Proposed Solution

0 1000 2000 3000 4000 5000 6000 7000 800010

20

30

40

50

60

70

80

90

100

110

Frequency (Hz)

Am

plit

ude

(d

B)

• 1) Do our best to recover the wideband information from narrowband speech

• 2) Use coding for the information that cannot be recovered

– Recovered information :• Low-frequency band

• High-frequency excitation

– Coded information :• High-frequency spectral

envelope

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 5

System Overview

• Inverse IRM filter is optional– produces a flat response from 200-3500 Hz

InverseIRMFilter

Low-frequencyregeneration

2

High-frequencyregeneration

8 kHznarrowband

16 kHzwideband

50-300 Hzband

300-3400 Hzband

3400-8000 HzbandSide information

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 6

Low-Frequency Regeneration (1/2)

• Assumptions :– Only pitch harmonics need to be recovered

• In general, no more than two pitch harmonics below 200 Hz

– Absolute phase is not perceptually relevant

• Frequency of harmonics determined from pitch analysis

• Amplitudes determined from feed-forward multi-layer perceptron (output in log domain)

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 7

Low-Frequency Regeneration (2/2)

Low-frequencyharmonic synthesis

Scaleand sum

21st harmonic

2nd harmonic

Pitchanalysis

MFCCcalculation

Multi-layerPerceptron

Pitch delay

Cepstral coefficients

Scalefactors

Narrowbandspeech

Lowfrequencies

Pitch gain

LPfilter

(1)

(1)

(16)

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 8

High-Frequency Extension

• Excitation-filter model (16 ms frames)

• Problem is separated in two parts– Excitation extension (no side information)– Spectral envelope coding (side information)

A(z)Narrowband

speech

LPCanalysis

Excitationextension B(z)

1

Spectral envelopeExtension

Side information(High-frequency spectral envelope)

High-frequency

band

Highpass

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 9

Excitation Extension

Narrowbandexcitation

Absolutevalue

Whiteningfilter

widebandexcitation

2 Highpass

0 5 10 15 20-0.5

0

0.5

1

0 2 4 6 80

2

4

6

8

10

0 5 10 15 200

0.2

0.4

0.6

0.8

1

0 2 4 6 80

5

10

15

0 5 10 15 20-0.2

0

0.2

0.4

0.6

0.8

0 2 4 6 80

1

2

3

4

5

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 10

Spectral Envelope Coding

• Spectral envelope calculated from the wideband LPC coefficients

• Quantization of the 3000-8000 Hz range (40 points)– Log domain– 8-bit Vector Quantization (500 bits/s side information,

using 16 ms frames)

• Concatenation with envelope obtained from LPC analysis on narrowband speech

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 11

Objective results

• Low-frequency band– 3 dB RMS error on harmonic amplitude

• High-frequency band– 3.6 dB RMS error on envelope– No objective measure for excitation extension

(perceptually very close to original)

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 12

Subjective Results

female male

Original wideband

Recovered fromoriginal IRM-filtered speech

Recovered fromG.729 coded speech

Speech Coding Workshop 2000 Jean-Marc Valin, Roch Lefebvre 13

Discussion

• Highlights– Expand IRM-filtered telephone-band speech to AM band– Very low side information rate (500 bits/s)

• Areas of improvement– Use high-band spectral estimation before coding– Use residual low-frequency information (below 300 Hz)– Noise robustness– Post-filtering