recording best practice to get best result ? about allosaurus

[allan-simon]

for example I have worked with people from https://shtooka.net to make recording of words so could I use them (as i can get the IPA of these recordings quite easily) to create a french model ?

[allan-simon]

Sorry i had not read the part about fine tuning , i will try to use the instruction to create a model for french

[xinjli]

hi, thanks for your question!

There is another model called interspeech21 which I have not included in the main branch.
It is in the pull request and should already be able to work after you merge that.
You can download that model and run inference by specifying that model, it should give you better results than the current default one.

Another thing you might do is the fine-tuning as you mentioned. It will significantly boost your results when even some small datasets.

[allan-simon]

ok thanks , so you would advice to use your branch + create a fine tune model ?

I continued thinking about it today and I was wondering something about my above use case :

If I create a model trained with french words only pronounced by native speakers , wouldn't that create a bias ?

i.e if one of my Chinese friends wants to correct their accent, allosaurus using that model will not recognize some mandarin-specific phone and will try hard instead to match it to the closest french phone (which is not good in my case because I want to point out their pronunciation is bad ) ? Instead shouldn't I train a mode with both mandarin+french , so that if they pronounce a french words with a strong chinese accent I will be able to correctly transcribe it ?

[xinjli]

Yeah, I think you can try using the new branch first and see whether it will work for not.
The new branch is trained with both Mandarin and French, so should be able to distinguish phones in both languages

If you fine-tune the model using native speakers, it might work very well for mandarin-speakers as you point out. Training with both languages might be one option.

[allan-simon]

I tried your branch but I'm running into the following issues

1. torchaudio was missing as dependencies from setup.py ( adding to the setup.py seems to be enough)

allospeech is not to be found

Traceback (most recent call last):
  File "/usr/lib/python3.8/runpy.py", line 194, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/usr/lib/python3.8/runpy.py", line 87, in _run_code
    exec(code, run_globals)
  File "/home/asimon/git/allosaurus/allosaurus/run.py", line 1, in <module>
    from allosaurus.app import read_recognizer
  File "/home/asimon/git/allosaurus/allosaurus/app.py", line 4, in <module>
    from allosaurus.pm.factory import read_pm
  File "/home/asimon/git/allosaurus/allosaurus/pm/factory.py", line 4, in <module>
    from allosaurus.utils.config import dotdict
  File "/home/asimon/git/allosaurus/allosaurus/utils/config.py", line 4, in <module>
    from allospeech.config import allospeech_config
ModuleNotFoundError: No module named 'allospeech'

[xinjli]

I see. allospeech is my private library I used to develop models.
I will fix them. sorry for the inconvenience.

[allan-simon]

ok no problem :) it's already very nice of you to both opensource the library and to take time to answer my questions

[xinjli]

yeah, thanks!