mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-09-08 08:34:19 +00:00
Code Issues Packages Projects Releases Wiki Activity

mgb2

Browse Source
This commit is contained in:
AmirHussein96 2022-06-05 01:00:32 +03:00
parent 86e1f9b056
commit 68aa924eeb
2 changed files with 61 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

51
egs/mgb2/ASR/README.md
View File

@ -1,30 +1,35 @@
# Introduction
# MGB2
Please refer to <https://icefall.readthedocs.io/en/latest/recipes/librispeech/index.html> for how to run models in this recipe.
The Multi-Dialect Broadcast News Arabic Speech Recognition (MGB-2):
The second edition of the Multi-Genre Broadcast (MGB-2) Challenge is
an evaluation of speech recognition and lightly supervised alignment
using TV recordings in Arabic. The speech data is broad and multi-genre,
spanning the whole range of TV output, and represents a challenging task for
speech technology. In 2016, the challenge featured two new Arabic tracks based
on TV data from Aljazeera. It was an official challenge at the 2016 IEEE
Workshop on Spoken Language Technology. The 1,200 hours MGB-2: from Aljazeera
TV programs have been manually captioned with no timing information.
QCRI Arabic ASR system has been used to recognize all programs. The ASR output
was used to align the manual captioning and produce speech segments for
training speech recognition. More than 20 hours from 2015 programs have been
transcribed verbatim and manually segmented. This data is split into a
development set of 10 hours, and a similar evaluation set of 10 hours.
Both the development and evaluation data have been released in the 2016 MGB
challenge
[./RESULTS.md](./RESULTS.md) contains the latest results.
Official reference:
# Transducers
Ali, Ahmed, et al. "The MGB-2 challenge: Arabic multi-dialect broadcast media recognition."
2016 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2016.
There are various folders containing the name `transducer` in this folder.
The following table lists the differences among them.
IEEE link: https://ieeexplore.ieee.org/abstract/document/7846277
| | Encoder | Decoder | Comment |
|---------------------------------------|---------------------|--------------------|---------------------------------------------------|
| `transducer` | Conformer | LSTM | |
| `transducer_stateless` | Conformer | Embedding + Conv1d | Using optimized_transducer from computing RNN-T loss |
| `transducer_stateless2` | Conformer | Embedding + Conv1d | Using torchaudio for computing RNN-T loss |
| `transducer_lstm` | LSTM | LSTM | |
| `transducer_stateless_multi_datasets` | Conformer | Embedding + Conv1d | Using data from GigaSpeech as extra training data |
| `pruned_transducer_stateless` | Conformer | Embedding + Conv1d | Using k2 pruned RNN-T loss |
| `pruned_transducer_stateless2` | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss |
| `pruned_transducer_stateless3` | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss + using GigaSpeech as extra training data |
| `pruned_transducer_stateless4` | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless2 + save averaged models periodically during training |
| `pruned_transducer_stateless5` | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + more layers + random combiner|
| `pruned_transducer_stateless6` | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + distillation with hubert|
| `pruned_stateless_emformer_rnnt2` | Emformer(from torchaudio) | Embedding + Conv1d | Using Emformer from torchaudio for streaming ASR|
## Performance Record (after 3 epochs)
| Decoding method | dev WER | test WER |
|---------------------------|------------|---------|
| attention-decoder | 27.87 | 26.12 |
| whole-lattice-rescoring | 25.32 | 23.53 |
The decoder in `transducer_stateless` is modified from the paper
[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
We place an additional Conv1d layer right after the input embedding layer.
See [RESULTS](/egs/mgb2/ASR/RESULTS.md) for details.

39
egs/mgb2/ASR/RESULTS.md
View File

@ -1,20 +1,33 @@
# Results
### MGB2 BPE training results (Conformer-CTC)
### MGB2 BPE training results (Conformer-CTC) (after 3 epochs)
#### 2022-06-04
The best WER, as of 2022-06-04, for the MGB2 test dataset is below
(using HLG decoding + n-gram LM rescoring + attention decoder rescoring):
| | dev | test |
Using whole lattice HLG decoding + n-gram LM rescoring + attention decoder rescoring
| | dev | test |
|-----|------------|------------|
| WER | - | - |
| WER | 25.32 | 23.53 |
Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
| ngram_lm_scale | attention_scale |
|----------------|-----------------|
| - | - |
| 0.1 | - |
Using n-best (n=0.5) HLG decoding + n-gram LM rescoring + attention decoder rescoring:
| | dev | test |
|-----|------------|------------|
| WER | 27.87 | 26.12 |
Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
| ngram_lm_scale | attention_scale |
|----------------|-----------------|
| 0.01 | 0.3 |
To reproduce the above result, use the following commands for training:
@ -40,7 +53,7 @@ export CUDA_VISIBLE_DEVICES="0,1"
```
and the following command for decoding
and the following command for nbest decoding
```
./conformer_ctc/decode.py \
@ -55,6 +68,20 @@ and the following command for decoding
--nbest-scale 0.5
```
and the following command for whole-lattice decoding
```
./conformer_ctc/decode.py \
--lang-dir data/lang_bpe_5000 \
--max-duration 30 \
--concatenate-cuts 0 \
--bucketing-sampler 1 \
--num-paths 1000 \
--epoch 2 \
--avg 2 \
--method whole-lattice-rescoring
```
You can find the pre-trained model by visiting
<comming soon>