Support torchscript.

[csukuangfj]

No description provided.

[danpovey]

It would be nice if it were more obvious from the binary itself what it does and how to use it.. or at least a pointer to where to find this info.

[danpovey]

mm, I see it has a usage message; is it accurate? (./bin/decode.py)... and what does this binary do?
I'm just wondering what kind of thing a jit file can contain, is it a nnet evaulation? With fixed size?

[csukuangfj]

This is just a prototype. I will add more documentation later.

The intention of this file is to replicate what we are doing in decode.py in icefall.

[csukuangfj]

I'm just wondering what kind of thing a jit file can contain, is it a nnet evaulation? With fixed size?

As far as I know, the torch scripted model behaves similarly to torch.nn.Module. You can vary the input shape, i.e.,
the N and T in (N, T, C).

Checkpoints just save the parameters of a model, without topology information.
If you want to recover from some checkpoint, you have to first build a nn.Module through some Python code,
which contains the definition of the model.

With a torch scripted model, all you need is a binary file ( I don't know what exactly is saved in it), load
it in C++, feed some input into it, and get the output for later processing.

[pzelasko]

the jit file has model weights and the model code translated into TorchScript intermediate representation (IR). AFAIK it doesn’t speed up the inference in any way unless the IR code is further optimized; e.g. it can be translated/exported to ONNX.

Also I think it is possible to extend both TorchScript and ONNX with custom ops (by building shared library extensions: https://pytorch.org/tutorials/advanced/torch_script_custom_ops) — maybe it makes sense to build k2 custom ops and involve LM + decoding as part of the JIT model for deployment packaging, I guess a lot of people would find it very convenient (same could be done for kaldifeat for a single-file model that outputs transcripts directly). Some food for thought.

[csukuangfj]

maybe it makes sense to build k2 custom ops and involve LM + decoding as part of the JIT model for deployment packaging,

Turns out custom operators supported by PyTorch has a few restrictions. i.e.,

The TorchScript compiler understands a fixed number of types. Only these types can be used as arguments to your custom operator. Currently these types are: torch::Tensor, torch::Scalar, double, int64_t and std::vector s of these types. Note that only double and not float, and only int64_t and not other integral types such as int, short or long are supported.

We need to pass FSAs to decoding functions, which I think is not supported by custom ops.

---

[EDITED]: See https://pytorch.org/tutorials/advanced/torch_script_custom_ops.html

[pzelasko]

Oh, that’s too bad. I guess it’s possible to convert an FSA to a bunch of tensors and reconstruct it inside but it’s probably too much hassle…

[csukuangfj]

Oh, a typo here. It should be ./bin/decode

If you clone k2 and do

mkdir build
cd build
cmake ..
make decode

you will find a binary in ./bin/decode

[csukuangfj]

Now it supports CTC decoding in C++ with a pre-trained model from icefall.
(Taking as inputs a file containing features, a file containing the scripted model, and a file containing a pre-trained BPE model,
it outputs the decoded transcripts).

$ ./bin/decode  --jit_pt ./cpu_jit.pt --feature_pt ./feature.pt  --bpe_model ./bpe.model
[I] /ceph-fj/fangjun/open-source/k2-torchscript/k2/torch/bin/decode.cc:200:int main(int, char**) THE GOOD NATURED AUDIENCE IN PITY TO FALLEN MAJESTY SHOWED FOR ONCE GREATER DEFERENCE TO THE KING THAN TO THE MINISTER AND SUNG THE PSALM WHICH THE FORMER HAD CALLED FOR THE OLD SERVANT TOLD HIM QUIETLY
AS THEY CREPT BACK TO GAMEWELL THAT THIS PASSAGEWAY LED FROM THE HUT AND THE PLEASANTS TO SHERWOOD AND THAT JEEOFFREY FOR THE TIME WAS HIDING WITH THE OUTLAWS IN THE FOREST BUT THE ESSENCE OF LUTHER'S LECTURES IS THERE

It will eventually take sound files as input and supports n-gram LM in decoding.

[danpovey]

Great!!

…

On Wed, Oct 6, 2021 at 8:09 PM Fangjun Kuang ***@***.***> wrote: Now it supports CTC decoding in C++ with a pre-trained model from icefall. (Taking as inputs a file containing features, a file containing the scripted model, and a file containing a pre-trained BPE model, it outputs the decoded transcripts). $ ./bin/decode --jit_pt ./cpu_jit.pt --feature_pt ./feature.pt --bpe_model ./bpe.model [I] /ceph-fj/fangjun/open-source/k2-torchscript/k2/torch/bin/decode.cc:200:int main(int, char**) THE GOOD NATURED AUDIENCE IN PITY TO FALLEN MAJESTY SHOWED FOR ONCE GREATER DEFERENCE TO THE KING THAN TO THE MINISTER AND SUNG THE PSALM WHICH THE FORMER HAD CALLED FOR THE OLD SERVANT TOLD HIM QUIETLY AS THEY CREPT BACK TO GAMEWELL THAT THIS PASSAGEWAY LED FROM THE HUT AND THE PLEASANTS TO SHERWOOD AND THAT JEEOFFREY FOR THE TIME WAS HIDING WITH THE OUTLAWS IN THE FOREST BUT THE ESSENCE OF LUTHER'S LECTURES IS THERE It will eventually take sound files as input and supports n-gram LM in decoding. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#839 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAZFLO7UIEIAFYCPSJ4WJMLUFQ4ATANCNFSM5FHUWOLQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[danpovey]

OK, so if I understand correctly, what is happening here is that you are using TorchScript to encode the model itself, then just coding the FSA-related parts in C++ directly?

And I believe the current status of this is that, since we haven't finished creating a C++ version of the Fsa object that we use in Python, things like automatic handling of attributes will not be happening right now, and these have to be coded directly?

BTW, it still does worry me a little, that the attribute handling will become much more opaque once we convert it to C++. But I also don't see another way to easily enable C++ deployment.

[csukuangfj]

The purpose of this pull-request is to show that we can import a pre-trained neural network model from PyTorch and
use it with k2 in C++.
(You can see that the code is not well organized and not well documented)

As for the attributes, this pull-request only handles aux_labels.

---

We will use k2::RaggedArc later, which handles attribute propagations automatically.
See https://github.com/k2-fsa/k2/blob/v2.0-pre/k2/python/csrc/torch/v2/ragged_arc.h

the attribute handling will become much more opaque once we convert it to C++

I don't think it is opaque. We can use it just like what we are doing in Python, i.e., c_fsa = a_fsa.IntersectDense(b_fsa)
will handle the attributes of c_fsa automatically.

k2::RaggedArc currently has Python dependencies so I am not using it in this pull-request.

Will remove Python dependencies from k2::RaggedArc later.

[danpovey]

Ah I see, cool.
I am wondering whether a different name might be better, than RaggedArc. Perhaps FsaClass?
Another possibility would be to rename Fsa,FsaVec,FsaOrVec to RawFsa, RawFsaVec, RawFsaOrVec, and rename RaggedArc to Fsa. That would be long-term a better long-term solution, IMO, because the names would match those in Python, but would likely require more messing about with documentation and so on.

[pkufool]

Will remove Python dependencies from k2::RaggedArc later.

@csukuangfj Have you started this stuff ? If not, I want to have a try. The plan is to replace the pyobject with torch::IValue.

[csukuangfj]

Will remove Python dependencies from k2::RaggedArc later.

@csukuangfj Have you started this stuff ? If not, I want to have a try. The plan is to replace the pyobject with torch::IValue.

Sure, you can go ahead.

[csukuangfj]

Now we can load HLG.pt in C++ to reconstruct a k2::Fsa. Will use it to decode.
And later we can load G.pt in C++ and use it for rescoring.

[csukuangfj]

Now it supports both CTC decoding and HLG decoding using pre-computed features.

Will implement decoding directly by reading sound files using https://github.com/csukuangfj/kaldifeat

[csukuangfj]

Now it takes sound files as input and outputs the recognition results.

Examples

CTC decoding

$ ./bin/decode \
--ue_ctc_decoding 1  \
--jit_pt ./cpu_jit.pt \
--bpe_model ./bpe.model \
./1089-134686-0000.wav  \
./1089-134686-0001.wav

[I] /ceph-fj/fangjun/open-source/k2-torchscript/build-cpu/k2/torch/bin/decode.cc:358:int main(int, char**)
Decoding result:

./1089-134686-0000.wav
HE HOPED THERE WOULD BE STEW FOR DINNER TURNIPS AND CARROTS AND BRUISED POTATOES AND FAT MUTTON PIECES TO BE LADLED OUT IN THICK PEPPERED FLOUR FAT AND SAUCE

./1089-134686-0001.wav
STUFFOR IT INTO YOU HIS BELLY COUNSELLED HIM

HLG decoding

$ ./bin/decode \
--use_ctc_decoding 0  \
--jit_pt ./cpu_jit.pt \
--hlg ./HLG.pt \
--word_table ./words.txt \
./1089-134686-0000.wav  \
./1089-134686-0001.wav

[I] /ceph-fj/fangjun/open-source/k2-torchscript/build-cpu/k2/torch/bin/decode.cc:358:int main(int, char**)
Decoding result:

./1089-134686-0000.wav
HE HOPED THERE WOULD BE STEW FOR DINNER TURNIPS AND CARROTS AND BRUISED POTATOES AND FAT MUTTON PIECES TO BE LADLED OUT IN THICK PEPPERED FLOUR FATTEN SAUCE

./1089-134686-0001.wav
STUFFED INTO YOU HIS BELLY COUNSELLED HIM

The ground truth corresponding to the sound files is:

[csukuangfj]

Will add GPU support, refactor the code, and add more documentation.

[csukuangfj]

Because of the issue mentioned in pytorch/pytorch#47493,
k2/torch/csrc/deserialization.cu causes compilation errors with NVCC when using torch 1.7.1

That issue has been fixed in https://github.com/pytorch/pytorch/pull/47492/files, which is
only available from torch 1.8.0.

I am afraid we have to use at least torch 1.8.0 for deployment.

[csukuangfj]

Here are some decoding logs of this pull request.

You can get an impression of the decoding speed by looking at the time stamp of the logs.

GPU + CTC decoding

$ CUDA_VISIBLE_DEVICES=2 ./bin/decode --use_gpu 1 --use_ctc_decoding 1 --jit_pt ./cpu_jit.pt --bpe_model ./bpe.model ./1089-134686-0000.wav ./1089-134686-0001.wav ./1089-134686-0002.wav ./1089-134686-0003.wav ./1089-134686
-0004.wav ./1089-134686-0005.wav ./1089-134686-0006.wav ./1089-134686-0007.wav ./1089-134686-0008.wav ./1089-134686-0009.wav
2021-10-12 00:04:18.597 [I] k2/torch/bin/decode.cu:121:int main(int, char**) Device: cuda:0
2021-10-12 00:04:23.648 [I] k2/torch/bin/decode.cu:140:int main(int, char**) Load wave files
2021-10-12 00:04:23.656 [I] k2/torch/bin/decode.cu:147:int main(int, char**) Compute features
2021-10-12 00:04:23.836 [I] k2/torch/bin/decode.cu:155:int main(int, char**) Load neural network model
2021-10-12 00:04:24.647 [I] k2/torch/bin/decode.cu:174:int main(int, char**) Compute nnet_output
2021-10-12 00:04:25.486 [I] k2/torch/bin/decode.cu:196:int main(int, char**) Build CTC topo
2021-10-12 00:04:25.488 [I] k2/torch/bin/decode.cu:208:int main(int, char**) Decoding
2021-10-12 00:04:25.629 [I] k2/torch/bin/decode.cu:254:int main(int, char**)
Decoding result:

./1089-134686-0000.wav
HE HOPED THERE WOULD BE STEW FOR DINNER TURNIPS AND CARROTS AND BRUISED POTATOES AND FAT MUTTON PIECES TO BE LADLED OUT IN THICK PEPPERED FLOUR FAT AND SAUCE

./1089-134686-0001.wav
STUFFOR IT INTO YOU HIS BELLY COUNSELLED HIM

./1089-134686-0002.wav
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS

./1089-134686-0003.wav
HELLO BERTI ANY GOOD IN YOUR MIND

./1089-134686-0004.wav
NUMBER TEN FRESH NELLIE IS WAITING ON YOU GOOD NIGHT HUSBAND

./1089-134686-0005.wav
THE MUSIC CAME NEARER AND HE RECALLED THE WORDS THE WORDS OF SHELLEY'S FRAGMENT UPON THE MOON WANDERING COMPANIONLESS PALE FOR WEARINESS

./1089-134686-0006.wav
THE DULL LIGHT FELL MORE FAINTLY UPON THE PAGE WHEREON ANOTHER EQUATION BEGAN TO UNFOLD ITSELF SLOWLY AND TO SPREAD ABROAD ITS WIDENING TAIL

./1089-134686-0007.wav
A COLD LUCID INDIFFERENCE REIGNED IN HIS SOUL

./1089-134686-0008.wav
THE CHAOS IN WHICH HIS ARDOUR EXTINGUISHED ITSELF WAS A COLD INDIFFERENT KNOWLEDGE OF HIMSELF

./1089-134686-0009.wav
AT MOST BY AN ALMS GIVEN TO A BEGGAR WHOSE BLESSING HE FLED FROM HE MIGHT HOPE WEARILY TO WIN FOR HIMSELF SOME MEASURE OF ACTUAL GRACE

CPU + CTC decoding

$ ./bin/decode --use_gpu 0 --use_ctc_decoding 1 --jit_pt ./cpu_jit.pt --bpe_model ./bpe.model ./1089-134686-0000.wav ./1089-134686-0001.wav ./1089-134686-0002.wav ./1089-134686-0003.wav ./1089-134686-0004.wav ./1089-134686
-0005.wav ./1089-134686-0006.wav ./1089-134686-0007.wav ./1089-134686-0008.wav ./1089-134686-0009.wav
2021-10-12 00:11:01.817 [I] k2/torch/bin/decode.cu:121:int main(int, char**) Device: cpu
2021-10-12 00:11:01.818 [I] k2/torch/bin/decode.cu:140:int main(int, char**) Load wave files
2021-10-12 00:11:01.825 [I] k2/torch/bin/decode.cu:147:int main(int, char**) Compute features
2021-10-12 00:11:01.946 [I] k2/torch/bin/decode.cu:155:int main(int, char**) Load neural network model
2021-10-12 00:11:02.449 [I] k2/torch/bin/decode.cu:174:int main(int, char**) Compute nnet_output
2021-10-12 00:11:17.42 [I] k2/torch/bin/decode.cu:196:int main(int, char**) Build CTC topo
2021-10-12 00:11:17.44 [I] k2/torch/bin/decode.cu:208:int main(int, char**) Decoding
2021-10-12 00:11:21.848 [I] k2/torch/bin/decode.cu:254:int main(int, char**)
Decoding result:

./1089-134686-0000.wav
HE HOPED THERE WOULD BE STEW FOR DINNER TURNIPS AND CARROTS AND BRUISED POTATOES AND FAT MUTTON PIECES TO BE LADLED OUT IN THICK PEPPERED FLOUR FAT AND SAUCE

./1089-134686-0001.wav
STUFFOR IT INTO YOU HIS BELLY COUNSELLED HIM

./1089-134686-0002.wav
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS

./1089-134686-0003.wav
HELLO BERTI ANY GOOD IN YOUR MIND

./1089-134686-0004.wav
NUMBER TEN FRESH NELLIE IS WAITING ON YOU GOOD NIGHT HUSBAND

./1089-134686-0005.wav
THE MUSIC CAME NEARER AND HE RECALLED THE WORDS THE WORDS OF SHELLEY'S FRAGMENT UPON THE MOON WANDERING COMPANIONLESS PALE FOR WEARINESS

./1089-134686-0006.wav
THE DULL LIGHT FELL MORE FAINTLY UPON THE PAGE WHEREON ANOTHER EQUATION BEGAN TO UNFOLD ITSELF SLOWLY AND TO SPREAD ABROAD ITS WIDENING TAIL

./1089-134686-0007.wav
A COLD LUCID INDIFFERENCE REIGNED IN HIS SOUL

./1089-134686-0008.wav
THE CHAOS IN WHICH HIS ARDOUR EXTINGUISHED ITSELF WAS A COLD INDIFFERENT KNOWLEDGE OF HIMSELF

./1089-134686-0009.wav
AT MOST BY AN ALMS GIVEN TO A BEGGAR WHOSE BLESSING HE FLED FROM HE MIGHT HOPE WEARILY TO WIN FOR HIMSELF SOME MEASURE OF ACTUAL GRACE

GPU + HLG decoding

$ CUDA_VISIBLE_DEVICES=2 ./bin/decode --use_gpu 1 --use_ctc_decoding 0 --jit_pt ./cpu_jit.pt --hlg ./HLG.pt --word_table ./words.txt ./1089-134686-0000.wav ./1089-134686-0001.wav ./1089-134686-0002.wav ./1089-134686-0003.w
av ./1089-134686-0004.wav ./1089-134686-0005.wav ./1089-134686-0006.wav ./1089-134686-0007.wav ./1089-134686-0008.wav ./1089-134686-0009.wav
2021-10-12 00:16:36.776 [I] k2/torch/bin/decode.cu:124:int main(int, char**) Device: cuda:0
2021-10-12 00:16:41.765 [I] k2/torch/bin/decode.cu:143:int main(int, char**) Load wave files
2021-10-12 00:16:41.771 [I] k2/torch/bin/decode.cu:150:int main(int, char**) Compute features
2021-10-12 00:16:41.778 [I] k2/torch/bin/decode.cu:158:int main(int, char**) Load neural network model
2021-10-12 00:16:42.424 [I] k2/torch/bin/decode.cu:177:int main(int, char**) Compute nnet_output
2021-10-12 00:16:43.77 [I] k2/torch/bin/decode.cu:204:int main(int, char**) Load HLG.pt
2021-10-12 00:16:46.867 [I] k2/torch/bin/decode.cu:212:int main(int, char**) Decoding
2021-10-12 00:16:47.194 [I] k2/torch/bin/decode.cu:258:int main(int, char**)
Decoding result:

./1089-134686-0000.wav
HE HOPED THERE WOULD BE STEW FOR DINNER TURNIPS AND CARROTS AND BRUISED POTATOES AND FAT MUTTON PIECES TO BE LADLED OUT IN THICK PEPPERED FLOUR FATTEN SAUCE

./1089-134686-0001.wav
STUFFED INTO YOU HIS BELLY COUNSELLED HIM

./1089-134686-0002.wav
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS

./1089-134686-0003.wav
HELLO BERTIE ANY GOOD IN YOUR MIND

./1089-134686-0004.wav
NUMBER TEN FRESH NELLIE IS WAITING ON YOU GOOD NIGHT HUSBAND

./1089-134686-0005.wav
THE MUSIC CAME NEARER AND HE RECALLED THE WORDS THE WORDS OF SHELLEY'S FRAGMENT UPON THE MOON WANDERING COMPANIONLESS PALE FOR WEARINESS

./1089-134686-0006.wav
THE DULL LIGHT FELL MORE FAINTLY UPON THE PAGE WHEREON ANOTHER EQUATION BEGAN TO UNFOLD ITSELF SLOWLY AND TO SPREAD ABROAD ITS WIDENING TAIL

./1089-134686-0007.wav
A COLD LUCID INDIFFERENCE REIGNED IN HIS SOUL

./1089-134686-0008.wav
THE CHAOS IN WHICH HIS ARDOUR EXTINGUISHED ITSELF WAS A COLD INDIFFERENT KNOWLEDGE OF HIMSELF

./1089-134686-0009.wav
AT MOST BY AN ALMS GIVEN TO A BEGGAR WHOSE BLESSING HE FLED FROM HE MIGHT HOPE WEARILY TO WIN FOR HIMSELF SOME MEASURE OF ACTUAL GRACE

CPU + HLG decoding

$ ./bin/decode --use_gpu 0 --use_ctc_decoding 0 --jit_pt ./cpu_jit.pt --hlg ./HLG.pt --word_table ./words.txt ./1089-134686-0000.wav ./1089-134686-0001.wav ./1089-134686-0002.wav ./1089-134686-0003.wav ./1089-134686-0004.w
av ./1089-134686-0005.wav ./1089-134686-0006.wav ./1089-134686-0007.wav ./1089-134686-0008.wav ./1089-134686-0009.wav
2021-10-12 00:19:19.795 [I] k2/torch/bin/decode.cu:124:int main(int, char**) Device: cpu
2021-10-12 00:19:19.796 [I] k2/torch/bin/decode.cu:143:int main(int, char**) Load wave files
2021-10-12 00:19:19.802 [I] k2/torch/bin/decode.cu:150:int main(int, char**) Compute features
2021-10-12 00:19:19.934 [I] k2/torch/bin/decode.cu:158:int main(int, char**) Load neural network model
2021-10-12 00:19:20.450 [I] k2/torch/bin/decode.cu:177:int main(int, char**) Compute nnet_output
2021-10-12 00:19:34.963 [I] k2/torch/bin/decode.cu:204:int main(int, char**) Load HLG.pt
2021-10-12 00:19:37.507 [I] k2/torch/bin/decode.cu:212:int main(int, char**) Decoding
2021-10-12 00:19:39.265 [I] k2/torch/bin/decode.cu:258:int main(int, char**)
Decoding result:

./1089-134686-0000.wav
HE HOPED THERE WOULD BE STEW FOR DINNER TURNIPS AND CARROTS AND BRUISED POTATOES AND FAT MUTTON PIECES TO BE LADLED OUT IN THICK PEPPERED FLOUR FATTEN SAUCE

./1089-134686-0001.wav
STUFFED INTO YOU HIS BELLY COUNSELLED HIM

./1089-134686-0002.wav
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS

./1089-134686-0003.wav
HELLO BERTIE ANY GOOD IN YOUR MIND

./1089-134686-0004.wav
NUMBER TEN FRESH NELLIE IS WAITING ON YOU GOOD NIGHT HUSBAND

./1089-134686-0005.wav
THE MUSIC CAME NEARER AND HE RECALLED THE WORDS THE WORDS OF SHELLEY'S FRAGMENT UPON THE MOON WANDERING COMPANIONLESS PALE FOR WEARINESS

./1089-134686-0006.wav
THE DULL LIGHT FELL MORE FAINTLY UPON THE PAGE WHEREON ANOTHER EQUATION BEGAN TO UNFOLD ITSELF SLOWLY AND TO SPREAD ABROAD ITS WIDENING TAIL

./1089-134686-0007.wav
A COLD LUCID INDIFFERENCE REIGNED IN HIS SOUL

./1089-134686-0008.wav
THE CHAOS IN WHICH HIS ARDOUR EXTINGUISHED ITSELF WAS A COLD INDIFFERENT KNOWLEDGE OF HIMSELF

./1089-134686-0009.wav
AT MOST BY AN ALMS GIVEN TO A BEGGAR WHOSE BLESSING HE FLED FROM HE MIGHT HOPE WEARILY TO WIN FOR HIMSELF SOME MEASURE OF ACTUAL GRACE

There are 10 test waves:

$ soxi *.wav

Input File     : '1089-134686-0000.wav'
Channels       : 1
Sample Rate    : 16000
Precision      : 16-bit
Duration       : 00:00:10.44 = 166960 samples ~ 782.625 CDDA sectors
File Size      : 334k
Bit Rate       : 256k
Sample Encoding: 16-bit Signed Integer PCM

Input File     : '1089-134686-0001.wav'                                                                                                     
Channels       : 1
Sample Rate    : 16000
Precision      : 16-bit
Duration       : 00:00:03.27 = 52400 samples ~ 245.625 CDDA sectors
File Size      : 105k
Bit Rate       : 256k
Sample Encoding: 16-bit Signed Integer PCM

Input File     : '1089-134686-0002.wav'
Channels       : 1
Sample Rate    : 16000
Precision      : 16-bit
Duration       : 00:00:06.62 = 106000 samples ~ 496.875 CDDA sectors
File Size      : 212k
Bit Rate       : 256k
Sample Encoding: 16-bit Signed Integer PCM

Input File     : '1089-134686-0003.wav'
Channels       : 1
Sample Rate    : 16000
Precision      : 16-bit
Duration       : 00:00:02.68 = 42880 samples ~ 201 CDDA sectors
File Size      : 85.8k
Bit Rate       : 256k
Sample Encoding: 16-bit Signed Integer PCM

Input File     : '1089-134686-0004.wav'                                                                                                     
Channels       : 1
Sample Rate    : 16000
Precision      : 16-bit
Duration       : 00:00:05.22 = 83441 samples ~ 391.13 CDDA sectors
File Size      : 167k
Bit Rate       : 256k
Sample Encoding: 16-bit Signed Integer PCM

Input File     : '1089-134686-0005.wav'
Channels       : 1
Sample Rate    : 16000
Precision      : 16-bit
Duration       : 00:00:09.63 = 154160 samples ~ 722.625 CDDA sectors
File Size      : 308k
Bit Rate       : 256k
Sample Encoding: 16-bit Signed Integer PCM

Input File     : '1089-134686-0006.wav'
Channels       : 1
Sample Rate    : 16000
Precision      : 16-bit
Duration       : 00:00:10.55 = 168880 samples ~ 791.625 CDDA sectors
File Size      : 338k
Bit Rate       : 256k
Sample Encoding: 16-bit Signed Integer PCM

Input File     : '1089-134686-0007.wav'
Channels       : 1
Sample Rate    : 16000
Precision      : 16-bit
Duration       : 00:00:04.28 = 68400 samples ~ 320.625 CDDA sectors
File Size      : 137k
Bit Rate       : 256k
Sample Encoding: 16-bit Signed Integer PCM

Input File     : '1089-134686-0008.wav'
Channels       : 1
Sample Rate    : 16000
Precision      : 16-bit
Duration       : 00:00:06.73 = 107680 samples ~ 504.75 CDDA sectors
File Size      : 215k
Bit Rate       : 256k
Sample Encoding: 16-bit Signed Integer PCM

Input File     : '1089-134686-0009.wav'
Channels       : 1
Sample Rate    : 16000
Precision      : 16-bit
Duration       : 00:00:10.57 = 169200 samples ~ 793.125 CDDA sectors
File Size      : 338k
Bit Rate       : 256k
Sample Encoding: 16-bit Signed Integer PCM

Total Duration of 10 files: 00:01:10.00

[csukuangfj]

Ready for review.

[csukuangfj]

I just wrote a Colab notebook to demonstrate how to use this pull request with a torch scripted model from icefall.

Please see
https://colab.research.google.com/drive/1BIGLWzS36isskMXHKcqC9ysN6pspYXs_?usp=sharing

Refer to the documentation in icefall https://icefall.readthedocs.io/en/latest/recipes/librispeech/conformer_ctc.html#deployment-with-c
for how to generate the required files. The above Colab notebook is downloading some pre-trained models
from hugging face.

[danpovey]

Cool!!

[csukuangfj]

Will merge it first so that @pkufool can continue his work.