-
Tested on Julia 1.8-rc1.
Pinning Zygote to 0.6.40 passes all tests.
```julia
(ApproximateGPs) pkg> test
Testing ApproximateGPs
┌ Warning: Could not use exact versions of packages in manifest…
-
I got an error from the dataloader when running the librispeech-100h stateless_transducer recipe after dumping features with LilcomChunkWriter, but this did not happen on the 1st epoch. The error see…
-
@csukuangfj Just wanted to note that the gradient is not incorrect for CPU vs GPU, the instructions clearly state that for CPU you need to provide log_softmax(joint-logits) whereas for the GPU you sho…
-
In the pretrained Transducer model, the greedy decoding strategy has a max_symbols_per_step parameter equal to 5. I was wondering why is it not "None" and shouldn't the transducer perform better witho…
-
I started installing K2, lhotse and Icefall. So far I was able to test K2 and it works perfectly, lhotse also works but when I tried to install icefall I got a weird issue about optimized-transducer. …
-
I started installing K2, lhotse and Icefall. So far I was able to test K2 and it works perfectly, lhotse also works but when I tried to install icefall I got a weird issue about optimized-transducer. …
-
i saw beta version of transducer loss relies on code from Alignment Restricted Streaming Recurrent Neural Network Transducer , maybe it's has parameters the left and right buffers bl/lr, how can i use…
-
I see there are two RNNT Loss. One from warp-transducer and the other from TF.
Which one do you recommend to use?
-
Hi . I saw that the pretrained Transducer model uses Numba implementation of RNNT-Loss. I was wondering what is the reason behind using numba instead of implementing it in native pytorch?
-
Hi, I am trying to use a Conformer Encoder as a Transducer Decoder in the VIVOS recipe:
```
# minibatch related
batch-size: 32r
maxlen-in: 800
maxlen-out: 150
# optimization related
criteri…