Evaluation procedure for audio watermark ver.5

Documents

CD Tracks 27, 32, 35, 40, 65, 66, 69, and 70
SQAM(SQAM_FLAC.zip; EBU TECHNICAL)
Freely available.
The initial and final segments of zero amplitude should be removed from the audio clip shown above. Their new initial and final samples are shown in the Table.

They should be used repetitively for a duration of 60 s.

New initial and final samples to clip from each SQAM track.
SQAM no.	initial	final	repetition is required
27	25390	726147	yes
32	24910	3145364	no
35	24904	2446802	yes
40	25302	2193536	yes
65	23937	4803885	no
66	22832	642775	yes
69	22888	1269672	yes
70	23365	733343	yes

Use the following data from the beginning bit. Error correction schemes can be used to embed the payload of 90 bits per 15 seconds.

payload.txtText file (CRLF)

# How to generate (Octave ver.3.x)
rand('state', 2012);
dat = randint(720,1);
save -mat payload.mat dat

Error rate may somewhat depend on the random seed used to randomize the start time of the stego signal for detection. Therefore, the results of error rates should be calculated as an arithmetic mean of 100 simulations with an arbitrary random seed and random start time.
The error rate of additive Gaussian noise may depend on a seed of randomize. The simulation should be done as the above manner.

PQevalAudio v2.0 (AFsp-v8r2.tar.gz; TSP lab. McGill Univ.)
PQevalAudio requires audio files to be 48-kHz sampling. The host signals must be converted from 44.1-kHz sampling to 48-kHz sampling using ResampAudio v5r1 which is included in AFsp-v8r2.
PQevalAudio can not evaluate degradation of single channel of sound quality for stereo signals. If the embedding is conducted for single channel, ODGs have to be obtained from monoaural signals of the signal channel to be embedded.

PICOLA(Courtesy of Prof. Mikio Ikeda, Yokkaichi University)
PICOLA is a tool for pitch-invariant time-scale conversion. Time-invariant frequency-scale conversion is achieved by combining PICOLA and sampling frequency conversion, ResampAudio.
The following is an example of -4 % time-invaliant frequency-scale conversion using the function file (available for Octave 3.0 or above).
```
 attacked_signal = pitch_change(stego_signal, 44100, 0.96);
```

The sampling frequency conversion attack requires expansion and contraction of duration and initial time for detection at the same rate of conversion.
For example, +10% conversion requires random selection from the initial 16.5 seconds samples (a 10% increase from the original) of the stego audio. Also, detection period is 49.5 seconds.

Encoding (96 kbps, HE profile)

neroAacEnc -he -br 96000 -if stego_wavfile.wav -of encoded_aacfile.mp4

Decoding

neroAacDec -if encoded_aacfile.mp4 -of decoded_wavfile.wav

Lame ver.3.99.3 lame-3.99.3.tar.gz
Encoding (joint-stereo, constant bit-rate、128 kbps)
```
lame --cbr -q 0 -m j -b 128
```

Decoding

lame mp3_file.mp3 --decode decoded_wavfile.wav

bpf.txtCoefficients of bandpass filter(text file)

# Example to use (Matlab or Octave)
load bpf.mat
attacked_signal = filter(bz, ap, stego_signal);

The following signal processing steps mimic attacks using DA and AD conversions.

Additive Gaussian noise at -80 dB (relative to maximum amplitude of 16-bit quantization as 0 dB)
Amplification of -2 dB to the above signal, followed by 16-bit quantization
Linear speed change conversion (pitch and time-scale conversion) of -0.1 %

da_ad_siml_revised.mCascade the above processing

# Example to use (Matlab or Octave)
attacked_signal = da_ad_siml_revised(stego_signal);