Not clear how to do a simple speech recognition

[foosoftsrlold]

It would be great if the instructions in the README were dumb-proof.

I just tried to follow them and the results were nonsensical.

It may clearly be due to error on our side or the environment (WSL) but looking at the release, I suspect that some data is missing (I just followed strictly what's on the README).

[Mte90]

Can you add more data about what steps dodn't worked for you?
We need more information to understand what is missing and what you are trying to do.

[foosoftsrlold]

What I'm trying to do is.... speech to text of a publicly available content
The system is Linux (WSL), bullsye/sid distribution

how to replicate:

virtualenv -p python3 venv
source venv/bin/activate
pip3 install deepspeech==0.8.0
wget https://radiomaria-cdn.thron.com/delivery/public/audio/radiomaria/75451e50-9ecd-482a-8201-08d1a9bf2189/hjtvzk/WEB/commento-al-messaggio-del-25-
ffmpeg -i commento-al-messaggio-del-25- -t 10 -ar 16000 commento-al-messaggio-del-25-.short.wav
deepspeech --model output_graph.pbmm --scorer scorer --audio commento-al-messaggio-del-25-.short.wav

The output of the "deepspeech" command is....

Loading model from file output_graph.pbmm
TensorFlow: v2.2.0-17-g0854bb5
DeepSpeech: v0.8.0-0-gf56b07d
2020-09-11 12:45:02.097968: I tensorflow/core/platform/cpu_feature_guard.cc:143] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
Loaded model in 0.0163s.
Loading scorer from files scorer
Loaded scorer in 0.00397s.
Running inference.
u vogue rire suu sur bert ahh non olh ful up uoh b
Inference took 6.754s for 10.000s audio file.

[Mte90]

Seems that you are using it right but you are not using our model for italian, right?

[foosoftsrlold]

Actually... I should be... I forgot these 2 lines which I had done

Scarica e scompatta i file per il modello italiano

curl -LO https://github.com/MozillaItalia/DeepSpeech-Italian-Model/releases/download/2020.08.07/model_tensorflow_it.tar.xz
tar xvf model_tensorflow_it.tar.xz

Just in case, this is the ls

(venv) luca@DESKTOP-QTQGTR0:/mnt/c/Users/luca/dati/voice/it$ ls -l
total 694012
-rwxrwxrwx 1 luca luca        86 Aug  7 00:36 alphabet.txt
-rwxrwxrwx 1 luca luca   9898126 Jun 26 10:51 commento-al-messaggio-del-25-
-rwxrwxrwx 1 luca luca    640078 Sep 11 12:40 commento-al-messaggio-del-25-.short.wav
-rwxrwxrwx 1 luca luca 315068264 Sep 11 12:54 model_tensorflow_it.tar.xz
-rwxrwxrwx 1 luca luca 188998006 Aug 10 11:51 output_graph.pbmm
-rwxrwxrwx 1 luca luca 196052112 Aug  7 00:41 scorer
drwxrwxrwx 1 luca luca       512 Sep 11 12:37 venv

[Mte90]

So you are using the model without transfer learning from english, that is pure italian but as we don't have a lot of hours to training is not very good.

So the procedure you are using is right, it is just the model that is not working good. I suggest to you to try the one with transfer learning from english that is more accurate as is adding over 7000+ hours compared to 250~ hours of italian (it is in the release page like the other one).

[nefastosaturo]

Hello there, I'm checking it right now. That mp3 file is a stereo file. During conversion, please convert it in mono.

Also you can try to use the same model but trained with transfer learning from english. You can find in the release or at this url:

https://github.com/MozillaItalia/DeepSpeech-Italian-Model/releases/download/2020.08.07/transfer_model_tensorflow_it.tar.xz

[nefastosaturo]

Ok it works. The disclaimer voice transcription:

Loaded model in 0.00849s.
Loading scorer from files /media/nefasto/data/deepspeech/models/ita/scorer
Loaded scorer in 0.000566s.
Running inference.
produzioni radi u eria tutti i diritti sono riservati
Inference took 1.648s for 3.971s audio file.

The first 10 seconds of speaker transcription (I cannot understand what he's saying btw)

Loaded scorer in 0.000662s.
Running inference.
cari mec le cime come tiamo e messaggio della regia della pace delle viticci cu mita denti festo ver creta nove su mali versario re ra palesin
Inference took 4.550s for 9.996s audio file.

[foosoftsrlold]

Ok! Way better with mono. A good candidate for README :)
Unfortunately the output... ehm... is not that fantastic. Any idea?

Original message is "Produzioni radio maria. Tutti i diritti sono riservati... blabla...".

Transfer learning model

(venv) luca@DESKTOP-QTQGTR0:/mnt/c/Users/luca/dati/voice/it$ deepspeech --model transfer_model_tensorflow_it/output_graph.pbmm --scorer scorer --audio commento-al-messaggio-del-25-.short.wav
...
...
produzioni radi umeri tutti i diritti sono riservati grave me nci di loci me come antiamo i mesca gio della gia dora pace delle veci mita dent

Normal model

(venv) luca@DESKTOP-QTQGTR0:/mnt/c/Users/luca/dati/voice/it$ deepspeech --model output_graph.pbmm --scorer scorer --audio commento-al-messaggio-del-25-.short.wav
...
...
zini radio eria puti diri ti sono riservati premerci li legione come diamo messaggio gelar giade la pace delle be dictu mido anti

[Mte90]

Readme updated, about the quality as today the model is not very good anyway.
We have some issues (find resources valid on legal purpose) about Italian datasets and also about the text corpora (that we built) used by the model.

If you want to help and contribute to the project I invite you to join us on Telegram with @mozitabot and later in the Developers group where we discuss.