Downloads · 30 days
0
csukuangfj/icefall-asr-librispeech-transducer-bpe-500-2021-12-17
icefall-asr-librispeech-transducer-bpe-500-2021-12-17 is a machine learning model from csukuangfj. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Catuion: You have to run git lfs pull. Otherwise, you will be SAD later.
Downloads · 30 days
0
Access
Public
Updated Dec 22, 2021
Repo size
376 MB
Likes
1
Public
Click a slice to open those files.
.pt376 MB · 99%
From the Hugging Face model README
sudo apt-get install git-lfs
git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-bpe-500-2021-12-17
cd icefall-asr-librispeech-transducer-bpe-500-2021-12-17
git lfs pull
Catuion: You have to run git lfs pull. Otherwise, you will be SAD later.
The model in this repo is trained using the commit cb04c8a7509425ab45fae888b0ca71bbbd23f0de.
You can use
git clone https://github.com/k2-fsa/icefall
cd icefall
git checkout cb04c8a7509425ab45fae888b0ca71bbbd23f0de
to download icefall.
You can find the model information by visiting https://github.com/k2-fsa/icefall/blob/cb04c8a7509425ab45fae888b0ca71bbbd23f0de/egs/librispeech/ASR/transducer/train.py#L196
In short, the encoder is a Conformer model with 8 heads, 12 encoder layers, 512-dim attention, 2048-dim feedforward; the decoder contains a 1024-dim embedding layer, plus a 4-layer LSTM with hidden size 512.
This repo provides pre-trained RNN-T Conformer model for the librispeech dataset using icefall.
The commands for training are:
cd egs/librispeech/ASR/
./prepare.sh
export CUDA_VISIBLE_DEVICES="0,1,2,3"
./transducer/train.py \
--world-size 4 \
--num-epochs 30 \
--start-epoch 0 \
--exp-dir transducer/exp-lr-2.5-full \
--full-libri 1 \
--max-duration 250 \
--lr-factor 2.5
The command for decoding is:
epoch=26
avg=12
./transducer/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir transducer/exp-lr-2.5-full \
--bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 100
You can find the decoding log for the above command in this repo: log/log-decode-epoch-26-avg-12-2021-12-17-09-33-04.
The best WER using greedy search is:
| test-clean | test-other | |
|---|---|---|
| WER | 3.16 | 7.71 |
preprained.ptexp/pretrained.pt is generated by the following command:
./transducer/export.py \
--epoch 26 \
--avg 12 \
--bpe-model data/lang_bpe_500/bpe.model \
--exp-dir transducer/exp-lr-2.5-full
HINT: To use pre-trained.pt to compute the WER for test-clean and test-other,
just do the following:
cp icefall-asr-librispeech-transducer-bpe-500-2021-12-17/exp/pretrained.pt \
/path/to/icefall/egs/librispeech/ASR/transducer/exp/epoch-999.pt
and pass --epoch 999 --avg 1 to transducer/decode.py.