FreeDV
/
LPCNet
mirror of https://github.com/drowe67/LPCNet.git
1
0
Fork 0
Code Issues Releases Wiki Activity

added direct-split VQ to process.sh - quite happy with the quality

pull/1/head
David 2019-03-22 14:13:23 +10:30
parent 660c8537cd
commit 8bd2c7e2e7
2 changed files with 12 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

34
process.sh
View File

@ -80,6 +80,7 @@ mkdir -p $PNG_PATH
rm -f $STATS
vq=pred_v2_stage
vq2=split_stage
# cp in originals
for f in $WAV_FILES
@ -128,7 +129,7 @@ if [ $lite -eq 0 ]; then
label=$(printf "33bit_20ms %-10s" "$f")
sox $WAVIN_PATH/$f.wav -t raw - | ./dump_data --test --c2pitch - - | \
./quant_feat -l "$label" -d 2 -o 6 -w --mbest 5 -q $vq'1.f32',$vq'2.f32',$vq'3.f32' -s $SV_PATH/$f'_5_33bit_20ms'.txt 2>>$STATS | \
./test_lpcnet - - | sox -r 16000 -t .s16 -c 1 - $WAVOUT_PATH/$f'_4_33bit_20ms'.wav
./test_lpcnet - - | sox -r 16000 -t .s16 -c 1 - $WAVOUT_PATH/$f'_5_33bit_20ms'.wav
done
# 33 bit 3 stage VQ searched with mbest algorithm, 30ms updates
@ -137,7 +138,7 @@ if [ $lite -eq 0 ]; then
label=$(printf "33bit_30ms %-10s" "$f")
sox $WAVIN_PATH/$f.wav -t raw - | ./dump_data --test --c2pitch - - | \
./quant_feat -l "$label" -d 3 -o 6 -w --mbest 5 -q $vq'1.f32',$vq'2.f32',$vq'3.f32' -s $SV_PATH/$f'_6_33bit_30ms'.txt 2>>$STATS | \
./test_lpcnet - - | sox -r 16000 -t .s16 -c 1 - $WAVOUT_PATH/$f'_5_33bit_30ms'.wav
./test_lpcnet - - | sox -r 16000 -t .s16 -c 1 - $WAVOUT_PATH/$f'_6_33bit_30ms'.wav
done
# 44 bit 4 stage VQ searched with mbest algorithm, 30ms updates
@ -146,20 +147,16 @@ if [ $lite -eq 0 ]; then
label=$(printf "44bit_30ms %-10s" "$f")
sox $WAVIN_PATH/$f.wav -t raw - | ./dump_data --test --c2pitch - - | \
./quant_feat -l "$label" -d 3 -o 6 -w --mbest 5 -q $vq'1.f32',$vq'2.f32',$vq'3.f32',$vq'4.f32' -s $SV_PATH/$f'_7_44bit_30ms'.txt 2>>$STATS | \
./test_lpcnet - - | sox -r 16000 -t .s16 -c 1 - $WAVOUT_PATH/$f'_6_44bit_30ms'.wav
./test_lpcnet - - | sox -r 16000 -t .s16 -c 1 - $WAVOUT_PATH/$f'_7_44bit_30ms'.wav
done
# Codec 2 and simulated SSB (10dB SNR, AWGN channel, 10Hz freq offset) reference samples
# non-predictive (direct) 44 bit 4 stage split VQ searched with mbest algorithm, 30ms updates
for f in $WAV_FILES
do
sox $WAVIN_PATH/$f.wav -t raw -r 8000 - | c2enc 2400 - - | c2dec 2400 - - | \
sox -r 8000 -t .s16 -c 1 - $WAVOUT_PATH/$f'_8_c2_2400'.wav
done
for f in $WAV_FILES
do
sox -r 16000 -c 1 $WAVIN_PATH/$f.wav -r 8000 -t raw - | \
cohpsk_ch - - -35 --Fs 8000 -f 10 --ssbfilt 1 | \
sox -r 8000 -t .s16 -c 1 - $WAVOUT_PATH/$f'_9_ssb_10dB'.wav
label=$(printf "44bit_sp_30ms %-10s" "$f")
sox $WAVIN_PATH/$f.wav -t raw - | ./dump_data --test --c2pitch - - | \
./quant_feat -l "$label" -d 3 -o 6 -i -p 0 --mbest 5 -q $vq2'1.f32',$vq2'2.f32',$vq2'3.f32',$vq2'4.f32' -s $SV_PATH/$f'_8_44bit_sp_30ms'.txt 2>>$STATS | \
./test_lpcnet - - | sox -r 16000 -t .s16 -c 1 - $WAVOUT_PATH/$f'_8_44bit_sp_30ms'.wav
done
fi # ... if [ $lite -eq 0 ] ...
@ -205,19 +202,12 @@ if [ $lite -eq 0 ]; then
<td>Same 33 bit VQ, but decimated down to 30ms rate, (33+8)/0.03 = 1367 bits/s</td>
</tr>
<tr>
<td>44bit_40ms</td>
<td>44bit_30ms</td>
<td>4 stage VQ, at 30ms update rate, (44+8)/0.03 = 1733 bits/s</td>
</tr>
<tr>
<td>c2_2400</td>
<td>Codec 2 at 2400 bits/s</td>
</tr>
<tr>
<td>ssb_10dB</td>
<td>Analog Single Sideband at a Signal/Noise ratio (in 3000 Hz) of 10dB. Band
limited to 300-2600Hz. At similar SNR with an appropriate modem we can send
error free 2000 bit/s coded speech.
</td>
<td>44bit_sp_30ms</td>
<td>Direct (non predictive) 4 stage split VQ, at 30ms update rate, (44+8)/0.03 = 1733 bits/s. First 3 stages are 18 elements wide, last stage is just 12. We quantise log magnitudes (Ly) rather than Ceptrals (dct(Ly)). Targeted at HF radio channel where predictive schemes perform poorly due to high bit error/packet error rate</td>
</tr>
</table>
<p>

4
src/quant_feat.c
View File

@ -314,10 +314,6 @@ int main(int argc, char *argv[]) {
for(i=0; i<NB_FEATURES; i++)
features_prev[dec][i] = features[i];
/*
for(i=0; i<NB_BANDS; i++)
features_mem[i] = 0.5*features_mem[i] + 0.5*features[i];
*/
// clear outpout features to make sure we are not cheating.
// Note we cant clear quant_out as we need memory of last
// frames output for pred quant