sherpa-onnx
sherpa-onnx copied to clipboard
Empty transcription with "parakeet-tdt-0.6b-v2" on some Files
Hi,
First of all, thank you very much for adding all the latest models to sherpa-onnx. However, it seems there might be an issue with the implementation of the recently added parakeet-tdt-0.6b-v2 model. Sherpa-onnx produces empty output for some files. I'm attaching an example from the GigaSpeech test set. While both the official implementation on Hugging Face and onnx_asr are able to produce results for this file, sherpa-onnx fails to do so.
Do you have any idea what might be causing this?
Example file download link: https://drive.google.com/file/d/1Ltd3xlS0-FEabO2dAlbOnpKR3XiItahm/view?usp=sharing
-
sherpa-onnx result: {"lang": "", "emotion": "", "event": "", "text": "", "timestamps": [], "tokens":[], "words": []} (using the model folder "sherpa-onnx-nemo-parakeet-tdt-0.6b-v2")
-
huggingface implementation result: Just before he passed away, when I visited him, he said, Howard, I'm working on becoming more of an optimal person. I'm just doing it now, and he was like so energetic. (https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2)
-
onnx-asr implementation result: Just before he passed away, when I visited him, he said, Howard, I'm working on becoming more of an optimal person. I'm just doing it now, and he was like so energetic. (https://github.com/istupakov/onnx-asr)
Thanks.
I am facing the same issue. I tried to investigate this a little bit and found a workaround by adding dithering.
Unfortunately, the internal dithering parameter is hardcoded to 0. (@csukuangfj, can we change this?)
So, you will have to add dithering to your input files directly.
It should have worked without dithering though, there is some issue in sherpa-onnx implementation or the model.
can we change this?
Yes, can you make a pull request to delete that line?
The default value of dither is still 0, so it won't change the current behavior.
You can change it using the command line by
--dither=0.001
to use 0.001 for dither.
@vsd-vector
Sorry, I don't understand. Did it work for you without removing the hardcoded value?
![]()
Sorry, I don't understand. Did it work for you without removing the hardcoded value?
No. It won't work.
I have to remove the hardco ded dither to make it work.