Accuracy issues on rpi

[JRMeyer]

>>> fotiDim
[December 14, 2017, 9:13pm]

Testing with the new binaries on RPi that
provided I am having accuracy
issues. I am testing with the following: slash
Binaries slash
Sample
model slash
2830-3980-0043.wav file which can be found
here

22.17.22.png.filename}[2286×148 35.3

I didn't try the exact same files on x86 Ubuntu but using the same audio
file with those model
files
gives accurate results.

[This is an archived TTS discussion thread from discourse.mozilla.org/t/accuracy-issues-on-rpi]

[JRMeyer]

>>> lissyx
[December 14, 2017, 9:18pm]

Wait, I did not noticed you were still using the output_graph.pb from
TaskCluster ; this is an overfitted model exclusively dedicated to test,
it has no purpose and it is actually expected to perform bad on anything
different than LDC93S1
![:slight_smile:](

[Archived Post]

[JRMeyer]

>>> fotiDim
[December 17, 2017, 9:31pm]

Is there a set of files that is supposed to work on the Pi then? Shall I
wait for the next release?

[Archived Post]

[JRMeyer]

>>> lissyx
[December 18, 2017, 2:52am]

They are on https://github.com/mozilla/DeepSpeech/releases/tag/v0.1.0

[Archived Post]

[JRMeyer]

>>> fotiDim
[December 18, 2017, 10:15am]

But these are the files I had the error with. Did you update them in the
meantime?

[Archived Post]

[JRMeyer]

>>> lissyx
[December 18, 2017, 1:08pm]

The error were related to the binaries published, not the model which
should be fine: output_graph.pb.

[Archived Post]

[JRMeyer]

>>> lissyx
[December 20, 2017, 11:25am]

[ slash fotiDim](
check that, please ? There may be multiple reasons for bad results, but
as long as you use our model and samples you should get good results.

[Archived Post]

[JRMeyer]

>>> fotiDim
[December 20, 2017, 11:32am]

I am fairly sure that are
exactly the files that I tried on the first place when you provided the
working binary. But just to double check I will give it another go
within the weekend so please bear with me until then.

[Archived Post]

[JRMeyer]

>>> lissyx
[December 20, 2017, 11:50am]

[ slash fotiDim](
re-tested, and I get the same result on my RPi3 and on my laptop on your
same sample: slash
On my laptop:

alexportable-alex:/tmp/deepspeech$ ./deepspeech models/output_graph.pb audio/2830-3980-0043.wav models/alphabet.txt -t
2017-12-20 12:43:52.266225: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
experience proves les
cpu_time_overall=8.36112 cpu_time_mfcc=0.00293 cpu_time_infer=8.35819

RPi3:

piraspberrypi:/models-release $ ~/native_client_arm/deepspeech models/output_graph.pb audio/2830-3980-0043.wav models/alphabet.txt -t
experience proves les
cpu_time_overall=82.25488 cpu_time_mfcc=0.02952 cpu_time_infer=82.22536

[Archived Post]

[JRMeyer]

>>> fotiDim
[January 15, 2018, 9:36pm]

sorry for the delay. Life
stroke in the meantime. So, there is apparently a difference with our
setups.

piraspberrypi:~/DeepSpeech/native_client_2 $ ./deepspeech ../models/output_graph.pb ../audio/2830-3980-0043.wav ../models/alphabet.txt -t
Killed

Also you are getting experience proves les. On my PC I was getting
experience proves this which happens to be the right text. Not sure
why you are getting inaccuracy.

Let me know if I can run another test or trace some logs somewhere.

[Archived Post]

[JRMeyer]

>>> lissyx
[January 16, 2018, 10:07am]

Don't feel sorry for the delay :/. But as you can see in my post above,
I don't have any problem on RPi3, so I don't know what is wrong on your
side. Can you ensure that:

[Archived Post]

[JRMeyer]

>>> fotiDim
[January 16, 2018, 9:30pm]

Do you use Raspbian or Raspbian Lite? Is it possible to upload your
image somewhere so I can give it a try?

[Archived Post]

[JRMeyer]

>>> lissyx
[January 16, 2018, 10:30pm]

697M free? That seems not that much. Check your PSU, even if it's
'specially made for RPi3', I got funny behaviors with some poor one.

It's Raspbian Jessie, but it should not have an impact:

piraspberrypi:~ $ lsb_release -a
No LSB modules are available.
Distributor ID: Raspbian
Description: Raspbian GNU/Linux 8.0 (jessie)
Release: 8.0
Codename: jessie

You latest output, with 'Killed' would suggest memory issue?

[Archived Post]

[JRMeyer]

>>> lissyx
[January 16, 2018, 10:33pm]

FTR:

piraspberrypi:~ $ cat /proc/meminfo
MemTotal: 994232 kB
MemFree: 377192 kB
MemAvailable: 904356 kB
Buffers: 68536 kB
Cached: 505380 kB
SwapCached: 1160 kB
Active: 287996 kB
Inactive: 291256 kB
Active(anon): 3112 kB
Inactive(anon): 12696 kB
Active(file): 284884 kB
Inactive(file): 278560 kB
Unevictable: 0 kB
Mlocked: 0 kB
SwapTotal: 102396 kB
SwapFree: 89380 kB
Dirty: 24 kB
Writeback: 0 kB
AnonPages: 4696 kB
Mapped: 11040 kB
Shmem: 10472 kB
Slab: 27556 kB
SReclaimable: 17548 kB
SUnreclaim: 10008 kB
KernelStack: 1000 kB
PageTables: 604 kB
NFS_Unstable: 0 kB
Bounce: 0 kB
WritebackTmp: 0 kB
CommitLimit: 599512 kB
Committed_AS: 76464 kB
VmallocTotal: 1064960 kB
VmallocUsed: 0 kB
VmallocChunk: 0 kB
CmaTotal: 8192 kB
CmaFree: 6796 kB
piraspberrypi:~ $

[Archived Post]