Automatic Speech Recognition
Transformers
Safetensors
English
lasr_ctc
medical-asr
radiology
medical

Output text from test_audio.wav contains many epsilons

#7
by robinchow - opened

When I run the sample pipeline code:

from transformers import pipeline
import huggingface_hub
from IPython.display import Audio, display
audio = huggingface_hub.hf_hub_download('google/medasr', 'test_audio.wav')
model_id = "google/medasr"
pipe = pipeline("automatic-speech-recognition", model=model_id)
result = pipe(audio,chunk_length_s=20,stride_length_s=2)
# the chunk length is how long in seconds MedASR batches audio and the stride length is the overlap between chunks.
print(result)

I get:

{'text': '<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> [ [EEXXAAMM<epsilon> T TYPPE]] C C<epsilon><epsilon>TT<epsilon><epsilon><epsilon><epsilon>  chchestest<epsilon><epsilon><epsilon><epsilon> P<epsilon><epsilon><epsilon>EE<epsilon><epsilon><epsilon><epsilon> pro pro<epsilon>ttoccolol<epsilon><epsilon> {periodperiod<epsilon>}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> [ [IINNDICCATIOON]]  5<epsilon><epsilon><epsilon>4<epsilon>--yyearar--oldold f feemmalee<epsilon>,<epsilon> sh shororttnessness of of b brereaathth,<epsilon><epsilon><epsilon> evvvalaluuateate<epsilon> for for<epsilon><epsilon> P P<epsilon>EE<epsilon><epsilon><epsilon> { {periodperiod}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> [TTEECHHNIIQUEE] Sttandardard<epsilon> pro prottooccol<epsilon><epsilon> { {periodperiod}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> [ [FFIINNDIINGGSS] { {coloncolon}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> P Pululmononarary  vaascuculalaturetureture<epsilon><epsilon> { {coloncoloncolon}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> The The<epsilon> ma mainin<epsilon><epsilon> P<epsilon><epsilon><epsilon>A<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> is is<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> pa patttentent<epsilon><epsilon><epsilon><epsilon><epsilon> { {periodperiod}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> There There<epsilon> are are<epsilon><epsilon><epsilon> f fillinging<epsilon> de de<epsilon>ffectectects<epsilon><epsilon><epsilon> in in<epsilon> the the<epsilon><epsilon><epsilon><epsilon><epsilon> se seggmentment<epsilon><epsilon>alal<epsilon> brranchcheses of of<epsilon><epsilon> the the<epsilon><epsilon><epsilon><epsilon> right right<epsilon><epsilon><epsilon><epsilon>  lowlowerer<epsilon><epsilon><epsilon> lo lobbee<epsilon><epsilon> { {commacommacomma}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> comp comp<epsilon><epsilon><epsilon>atat<epsilon>ibleible<epsilon><epsilon> with with<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> a a<epsilon>cucutete<epsilon><epsilon><epsilon> P<epsilon><epsilon>EE<epsilon><epsilon><epsilon> { {periodperiod}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> No No<epsilon><epsilon><epsilon> sadd<epsilon>ddlele<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> e embmbmbolusus<epsilon><epsilon><epsilon> { {periodperiod}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> L Luungngss<epsilon><epsilon> { {coloncoloncolon}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> No No<epsilon><epsilon> p pneneummooththororaaaxx<epsilon><epsilon><epsilon> { {periodperiodperiod}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> S Smmalll<epsilon><epsilon> b bililaaterteralalal<epsilon> e effffffususionionions<epsilon><epsilon><epsilon><epsilon><epsilon> { {commacomma}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> right right<epsilon><epsilon><epsilon><epsilon> great great great<epsilon>erer<epsilon> than than<epsilon><epsilon><epsilon> left left<epsilon><epsilon><epsilon><epsilon> { {periodperiodperiod}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> { {newnew<epsilon>  <epsilon><epsilon>paragraphparagraph}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> [ [IIMPPREES<epsilon>SSIOONN]]<epsilon><epsilon><epsilon> { {coloncolon}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon> A<epsilon><epsilon>cucutete<epsilon><epsilon> se se<epsilon>gmentmentment<epsilon>alal<epsilon><epsilon> P P<epsilon>EE<epsilon><epsilon><epsilon><epsilon> right right<epsilon><epsilon><epsilon>  lowlowerer<epsilon> lo lobbee<epsilon><epsilon><epsilon><epsilon> { {periodperiod}}<epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon><epsilon></s>'}

If you can provide some guidance as to what I need to look into to fix this, that'd be so appreciated.

  • Robin

Ever figure this out? I get epsilon and duplicated text, like

...<epsilon><epsilon><epsilon> on on<epsilon> a<epsilon> co couupple<epsilon> of of<epsilon><epsilon> high high<epsilon><epsilon><epsilon> b bloloodd p pressressururee<epsilon>...

Edit: It looks like something in current versions of transformers or another dependency breaks MedAsr. Their GoogleCollab example only works because they have transformers pinned to a specific git commit from last year.

I opened a GItHub issue, so maybe someone will notice. https://github.com/Google-Health/medasr/issues/12

Google org

Hi @robinchow
Before running the code, Can you please install the transformers library directly from the specific commit mentioned in the model card . I've found this version works without any issue .
$ uv pip install git+https://github.com/huggingface/transformers.git@65dc261512cbdb1ee72b88ae5b222f2605aad8e5

Please let me know if you face any issue .
Thanks

@pannaga10 I mention in the linked github issue above that that specific git commit seems to be the only one that works. The issue is that commit isn't any actual tagged release of transformers, and "use this old commit only, and never update transformers again" isn't really a great solution. I mention in the github issue I opened that rolling transformers back to that commit also backdates a bunch of other dependencies. It becomes a bit of a mess to try to manage all the different dependency versions to pick up other fixes and improvements in other libraries.


Edit

After writing this reply, I see you also commented in my GH issue. I appreciate you escalating the issue. Thanks for the help!

Sign up or log in to comment