---

Speech Codec Wav Samples
Keywords: Voice Codec Samples, Codec Sample Data, WAV File Samples

Overview | Speech Recognition | Speech Codec Samples | Speech + Noise Codec Samples | ITU ◳ | MELPe Speech Codecs ◳ | PESQ Speech Quality Measurement | AT&T Noise Preprocessor | Notes | More Samples | EVS Codec ◳ | Codec PC Simulation Program ◳

Overview

Below are a variety of "before and after" .wav file samples for different LBR (low bit rate) speech (voice) codecs, including MELP, GSM, and G.729A/B, with bit rates ranging from 600 bps to 13000 bps. Click on .wav file links in the table below to listen to the samples (note -- all .wav file entries are mono, sampled at 8 kHz).

Speech samples are at left, with different codec types across (columns). Each row is a different language or sample type, such as addition of background noise. In several cases, a language sample may include both male and female speakers.

Underneath each sample is given the PESQ score, which is a numerical algorithm comparison between the original sample and the processed sample designed to closely approximate a MOS score. 4.5 is a perfect PESQ score, meaning there was no degradation of the processed sample from the original sample. PESQ scores normally refer to the "Original" sample at far left column, unless otherwise indicated. More information on PESQ is given below

Speech Recognition

Signalogic uses these wav files in speech recognition training, testing, and analysis work, for example comparing noise reduction and silence detection algorithms utilized by state-of-the-art codecs with those used by popular speech recognition open source, such as Kaldi ◳. The most advanced codec in use today is the 3GPP sponsored EVS codec ◳, which accurately classifies and models voice, music and other sounds, and background noise, applies precision silence detection, and offers a wide variety of sampling rates and bitrates.

For Kaldi online decoding, the Kaldi ASR Offloading project offers a production alternative to GStreamer. SigSRF software, deployed worldwide by telecom, LEA, and analytics users, has an inference library that interfaces to Kaldi run-time libraries. This allows RTP packet audio streams using wideband codecs to be processed with full RFC and jitter buffer support, including multiple stream groups, then given to Kaldi's online decoding raw audio interface.

More information on Signalogic speech recognition projects can be found at:

• Kaldi Online Decoding RTP Packet Interface ◳
• Edge AI Server (TI Wiki "Small AI Server" page) ◳
• DeepDome™ Home/Business Assistant Privacy, with Two-Level Wake Word Security ◳
• SigSRF SDK Overview ◳
• DirectCore® Software ◳
• 2-D Spectrograph page ◳

Speech Codec Samples

Numbers given in () below are bitrate (in bps) for uncompressed wav file samples, and inherent algorithm frame size (in msec) for compressed (i.e. codec encoder output) wav file samples. Typically, the frame size is also the delay of the codec, but in some cases there may be additional "look ahead" delay.
  

Original Sample Fs = 8 kHz (128000 bps)

MELP² 2400 bps (22.5 msec)

MELPe 2400 bps (22.5 msec)

MELPe 1200 bps (67.5 msec)

MELPe-Plus 2400 bps (22.5 msec)

MELPe-Plus 2700 bps (20 msec)

MELPe-Plus 4000 bps (20 msec)

MELPe-Plus17 600 bps (30 msec)

G.729A 8000 bps (10 msec)

GSM³ 13000 bps (20 msec)

MELPe 2400 + AT&T NPP4

G.729A 8000 bps + AT&T NPP

GSM 13000 bps +AT&T NPP

CVSD 13000 bps +AT&T NPP

G726 13000 bps +AT&T NPP

Language & Score

Male

Female

Male

Female

Male

Female

Male

Female

Male

Female

Male

Female

Male

Female

Male

Female

Male

Female

Male

Female

Male

Female

Male

Female

Male

Female

Male

Female

Male

Female

English1

eng2_m

eng2_f

male600

female600

eng_m

eng_f

eng_m

eng_f

(ITU) PESQ Score

4.5

4.5

2.673

2.293

English2

eng_m

eng_f

eng_m1

eng_f1

eng_m2

eng_f2

eng_m3

eng_f3

eng_m4

eng_f4

eng_m5

eng_f5

eng_m6

eng_f6

eng_m7

eng_f7

eng_m9

eng_f9

eng_m10

eng_f10

(ITU) PESQ1 Score

4.5

4.5

2.666

2.445

2.86

2.583

2.413

2.323

2.923

2.704

2.958

2.734

3.266

3.078

3.570

3.265

2.583

2.434

3.254

3.076

French

f_m

f_f

f_m1

f_f1

f_m2

f_f2

f_m3

f_f3

f_m4

f_f4

f_m5

f_f5

f_m6

f_f6

f_m7

f_f7

f_m9

f_f9

f_m10

f_f10

f_m

f_f

f_m

f_f

(ITU) PESQ Score

4.5

4.5

2.401

2.549

2.482

2.575

2.249

2.365

2.786

2.756

2.770

2.829

3.162

3.243

3.349

3.352

2.482

2.599

3.343

3.315

German

(ITU) PESQ Score

Japanese

(ITU) PESQ Score

Chinese

ch_m

ch_f

ch_m1

ch_f1

ch_m2

ch_f2

ch_m3

ch_f3

ch_m4

ch_f4

ch_m5

ch_f5

ch_m6

ch_f6

ch_m7

ch_f7

ch_m9

ch_f9

ch_m10

ch_f10

(ITU) PESQ Score

4.5

4.5

2.781

2.572

3.080

2.769

2.738

2.477

3.120

2.739

3.124

2.809

3.490

3.164

3.730

3.601

2.969

2.641

3.548

3.462

NSA test vector5

nsa_m

nsa_f

nsa_m1

nsa_f1

nsa_m2

nsa_f2

nsa_m3

nsa_f3

nsa_m4

nsa_f4

nsa_m5

nsa_f5

nsa_m6

nsa_f6

nsa_m600

nsa_f600

nsa_m7

nsa_f7

nsa_m9

nsa_f9

nsa_m10

nsa_f10

(ITU) PESQ Score

4.5

4.5

3.185

2.963

3.270

3.063

2.976

2.761

3.331

3.029

3.330

2.963

3.637

3.451

2.694

2.275

3.882

3.901

3.197

2.988

3.865

3.659

GSM test vector

(ITU) PESQ Score

G.729A test vector

(ITU) PESQ Score

Speech + Noise Codec Samples

Original Sample w/o Noise Fs = 8 kHz (128000 bps)

Original Sample with Noise Fs = 8 kHz (128000 bps)

MELP² 2400 bps (22.5 msec)

MELPe 2400 bps (22.5 msec)

MELPe 1200 bps (67.5 msec)

MELPe-Plus 2400 bps (22.5 msec)

MELPe-Plus 2700 bps (20 msec)

MELPe-Plus 4000 bps (20 msec)

G.729A 8000 bps (10 msec)

GSM³ (20 msec) 13000 bps

RLS

AT&T NPP4

MELPe 2400 + AT&T NPP

G.729A + AT&T NPP 8000 bps

GSM +AT&T NPP

MELPe + RLS6

Language & Score

English + car background noise (1)

car

car1

car2

car3

car4

car5

car6

car7

car10

car9

car12

PESQ Score

2.458

2.561

2.37

2.740

2.723

2.968

3.51511

2.8367

3.65512

English + white noise (2)

white noise original

white noise

white noise1

white noise2

white noise3

white noise4

white noise5

white noise6

white noise7

white noise9

white noise10

white noise9

white noise11

white noise13

PESQ Score

4.5

1.615

2.251

2.237

1.574

2.509

2.795

2.933

3.5508

1.5139

2.94610

3.1219

English + wideband noise (3)

wideband noise original

wideband noise

wideband noise1

wideband noise2

wideband noise3

wideband noise4

wideband noise5

wideband noise6

wideband noise7

wideband noise9

wideband noise10

wideband noise11

wideband noise12

wideband noise14

PESQ Score

4.5

1.695

2.64313

2.69613

2.45313

2.85513

2.78813

3.15413

3.41813

3.913

2.08014

1.98914

2.06414

3.253

English + street background noise (4)

street noise original

street noise

street noise1

street noise2

street noise3

street noise4

street noise5

street noise6

street noise7

street noise9

street noise10

street noise11

street noise12

street noise14

PESQ Score

4.5

1.800

2.28315

2.35115

2.19715

2.49315

2.42215

2.92515

3.35915

3.799

2.13916

1.98916

2.13316

3.239

External Links -- More Samples

Below are listed more web pages with audio/voice codec samples:

http://www.kyastem.co.jp/english/e-sample.html

http://www.hawksoft.com/hawkvoice/codecs.shtml

SigSRF SDK (more wav files are included in the demo download) ◳

About PESQ

PESQ (Perceptual Evaluation of Speech Quality) is an ITU algorithm that is increasingly used to emulate MOS, which is "mean opinion score", a listening test for speech codecs. The objective is to automate and ensure reproducibility of measurement of degradation over telephony channels, due to speech compression, line conditions, delay/echo, etc. and obtain results that closely correlate with MOS human listening tests.

More information on the PESQ approach and algorithm can be found at: http://www.pesq.org/

The current ITU PESQ recommendation is P.862E ◳, and the current software version of ITU PESQ is v1.2, which is a significant improvement over earlier versions. There is still ongoing work by PESQ developers to address limitations for maximum input file size, worst-case variation in delay (dynamic time-alignment with reference input), and noisy backgrounds.

AT&T Noise Preprocessor

The AT&T website contains no published information about the AT&T Noise Preprocessor. Instead, below we provide the abstract of the AT&T patent:

The system and method of the invention relates to voice detection technology for determining instants of time at which a snapshot of noise characteristics results in improved adaptation of noise floors used in voice detection. The approach is based on the "lower envelope" of the smoothed input signal power. Incorporation of this approach in a simple time domain VAD (Voice Activity Detector) results in an effective low-complexity system which, on the basis of simulations, gives good performance down to SNR values of about 0 dB. In the invention the lower envelope also provides the updated value of the noise threshold during the presence of speech. The invention can also be embedded in other, more complex (e.g., frequency domain) VADs at low computational cost.

More information on the AT&T patent:

Patent Number	5,991,718
Author	David Malah
Title	System and method for noise threshold adaptation for voice activity detection in nonstationary noise environments

More information about AT&T Noise Preprocessor can be found in this 2001 ICASSP paper ◳.

Send e-mail to: Joe Alfred [ALIPM] at AT&T

Notes

¹ ITU P.862 v1.2. For more information, see About PESQ section above.
² Mixed Excitation Linear Predictive.
³ Global System for Mobile Communications.
⁴ AT&T Noise Preprocessor algorithm. NPP algorithm has some processing delay before algorithm output stabilizes.
⁵ Mixed Excitation Linear Predictive. MELP = original MELP v1.2, MELPe = enhanced MELP (current standard).
⁶ MELPe + RLS based adaptive noise cancellation information is located at http://www.owlnet.rice.edu/~ryanking/elec431/compare.html
⁷ PESQ score referenced to car10 waveform.
⁸ PESQ score referenced to white noise waveform.
⁹ PESQ score referenced to white noise original waveform.
¹⁰ PESQ score referenced to white_noise10 waveform.
¹¹ PESQ score referenced to car
¹² PESQ score referenced to car10
¹³ PESQ score referenced to wideband noise
¹⁴ PESQ score referenced to wideband noise original
¹⁵ PESQ score referenced to streetnoise
¹⁶ PESQ score referenced to streetnoise original
¹⁷ Preliminary Simulation of 600bps MELPe-Plus Codec.