Add dns_icassp22 Speech Enhancement Recipe #4657
Conversation
|
Could you format the Python scripts with
|
Just want to make sure. So I need to run these from the root directory, through all pieces of code in the package? |
To be precise, you need to go to the root directory of espnet and run these commands. |
|
@Emrys365 Hi, I've run |
Codecov Report
@@ Coverage Diff @@
## master #4657 +/- ##
==========================================
- Coverage 83.09% 83.09% -0.01%
==========================================
Files 518 518
Lines 44777 44700 -77
==========================================
- Hits 37207 37143 -64
+ Misses 7570 7557 -13
Flags with carried forward coverage won't be shown. Click here to find out more.
📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
egs2/TEMPLATE/asr1/db.sh
Outdated
| @@ -17,6 +17,7 @@ DIRHA_WSJ_PROCESSED="${PWD}/data/local/dirha_wsj_processed" # Output file path | |||
| DNS= | |||
| DNS2= | |||
| DNS3= | |||
| DNS4="${PWD}/data_dns4_raw" | |||
There was a problem hiding this comment.
It's better to use DNS4 or downloads here to follow the convention instead of data_dns4_raw
There was a problem hiding this comment.
Ok! will change to downloads
| - pytorch version: `pytorch 1.12.1+cu102` | ||
| - Git hash: `13db69d3befc3c82a5ff5a11e28bf79d5030603f` | ||
| - Commit date: `Mon Aug 29 13:44:35 2022 +0000` | ||
|
|
There was a problem hiding this comment.
how about uploading the model and add it here?
There was a problem hiding this comment.
Sure. I'm waiting for the model to train on 48kHz, full dataset (which might take several days more)
I guess you're talking about btw, in this case, the formatting won't conform to |
|
@slSeanWU If the file is not in this repo, formatting would not be a problem. |
|
|
||
| # get noisy wav synthesizer config file | ||
| cd ${PATH_PREFIX} | ||
| wget https://raw.githubusercontent.com/microsoft/DNS-Challenge/master/noisyspeech_synthesizer.cfg -O noisyspeech_synthesizer.cfg |
There was a problem hiding this comment.
Thanks for the action. Sorry I didn't make it clear. Can you use git clone and specify the commit hash information? Then use the file from cloned repo. This is just to avoid the possible incompatibility due to the future updates.
There was a problem hiding this comment.
This way, does it mean that we must clone the entire repo rather than just getting the files we need?
Or, is there a way we can just fetch those specific files? Thanks a lot.
There was a problem hiding this comment.
You could clone only a specific branch by
git clone -b [branch_name] ...| cd DNS-Challenge | ||
| git checkout 5582dcf5ba43155621de72a035eb54a7d233af14 | ||
| cp noisyspeech_synthesizer.cfg ../ | ||
| cp audiolib.py $RUN_DIR/local/ | ||
| cp utils.py $RUN_DIR/local/ | ||
| echo "Required scripts & files fetched from DNS-Challenge repo" | ||
|
|
There was a problem hiding this comment.
I use these to fetch the files now, do they look okay to you?
| @@ -26,4 +26,5 @@ test_sets="cv_synthetic tt_synthetic" | |||
| --use_noise_ref false \ | |||
| --max_wav_duration 31 \ | |||
| --inference_model "valid.loss.best.pth" \ | |||
| --nj 8 \ | |||
There was a problem hiding this comment.
Perhaps you could delete this line.
There was a problem hiding this comment.
I added this line since it's much easier to run into OOM error with the default nj=32 when we're working with 48 kHz.
There was a problem hiding this comment.
I mean in the preprocessing stages
|
Thanks, @slSeanWU! |
Hello,
Two things to note:
(1) I checked how DNS challenges evaluate submissions. They use subjective MOS on recorded test samples that don't come with a "clean reference", so I think it'll be hard for us to compare with those submitted implementations.
(2) The original script from DNS repo to synthesize training examples in single-threaded, hence very slow. I refactored it using
multiprocessingto enable parallelization.Thanks in advance for reviews & comments,
Shih-Lun