GigaSpeech recipe (k2-fsa#120)

* initial commit

* support download, data prep, and fbank

* on-the-fly feature extraction by default

* support BPE based lang

* support HLG for BPE

* small fix

* small fix

* chunked feature extraction by default

* Compute features for GigaSpeech by splitting the manifest.

* Fixes after review.

* Split manifests into 2000 pieces.

* set audio duration mismatch tolerance to 0.01

* small fix

* add conformer training recipe

* Add conformer.py without pre-commit checking

* lazy loading and use SingleCutSampler

* DynamicBucketingSampler

* use KaldifeatFbank to compute fbank for musan

* use pretrained language model and lexicon

* use 3gram to decode, 4gram to rescore

* Add decode.py

* Update .flake8

* Delete compute_fbank_gigaspeech.py

* Use BucketingSampler for valid and test dataloader

* Update params in train.py

* Use bpe_500

* update params in decode.py

* Decrease num_paths while CUDA OOM

* Added README

* Update RESULTS

* black

* Decrease num_paths while CUDA OOM

* Decode with post-processing

* Update results

* Remove lazy_load option

* Use default `storage_type`

* Keep the original tolerance

* Use split-lazy

* black

* Update pretrained model

Co-authored-by: Fangjun Kuang <[email protected]>

Author: wgb14

.flake8

@@ -7,6 +7,7 @@ per-file-ignores =
     egs/librispeech/ASR/*/conformer.py: E501,
     egs/aishell/ASR/*/conformer.py: E501,
     egs/tedlium3/ASR/*/conformer.py: E501,
+    egs/gigaspeech/ASR/*/conformer.py: E501,
     egs/librispeech/ASR/pruned_transducer_stateless2/*.py: E501,
 
     # invalid escape sequence (cause by tex formular), W605

.gitignore

@@ -6,6 +6,8 @@ exp
 exp*/
 *.pt
 download
+dask-worker-space
+log
 *.bak
 *-bak
 *bak.py

egs/gigaspeech/ASR/.gitignore

@@ -0,0 +1 @@
+log-*

egs/gigaspeech/ASR/README.md

@@ -0,0 +1,20 @@
+# GigaSpeech
+GigaSpeech, an evolving, multi-domain English
+speech recognition corpus with 10,000 hours of high quality labeled
+audio, collected from audiobooks, podcasts
+and YouTube, covering both read and spontaneous speaking styles,
+and a variety of topics, such as arts, science, sports, etc. More details can be found: https://github.com/SpeechColab/GigaSpeech
+
+## Download
+
+Apply for the download credentials and download the dataset by following https://github.com/SpeechColab/GigaSpeech#download. Then create a symlink
+```bash
+ln -sfv /path/to/GigaSpeech download/GigaSpeech
+```
+
+## Performance Record
+|     |  Dev  | Test  |
+|-----|-------|-------|
+| WER | 10.47 | 10.58 |
+
+See [RESULTS](/egs/gigaspeech/ASR/RESULTS.md) for details.

egs/gigaspeech/ASR/RESULTS.md

@@ -0,0 +1,79 @@
+## Results
+
+### GigaSpeech BPE training results (Conformer-CTC)
+
+#### 2022-04-06
+
+The best WER, as of 2022-04-06, for the gigaspeech is below
+
+Results using HLG decoding + n-gram LM rescoring + attention decoder rescoring:
+
+|     |  Dev  | Test  |
+|-----|-------|-------|
+| WER | 10.47 | 10.58 |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+|      0.5       |       1.3       |
+
+
+To reproduce the above result, use the following commands for training:
+
+```
+cd egs/gigaspeech/ASR
+./prepare.sh
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+./conformer_ctc/train.py \
+  --max-duration 120 \
+  --num-workers 1 \
+  --world-size 8 \
+  --exp-dir conformer_ctc/exp_500 \
+  --lang-dir data/lang_bpe_500
+```
+
+and the following command for decoding:
+
+```
+./conformer_ctc/decode.py \
+  --epoch 18 \
+  --avg 6 \
+  --method attention-decoder \
+  --num-paths 1000 \
+  --exp-dir conformer_ctc/exp_500 \
+  --lang-dir data/lang_bpe_500 \
+  --max-duration 20 \
+  --num-workers 1
+```
+
+Results using HLG decoding + whole lattice rescoring:
+
+|     |  Dev  | Test  |
+|-----|-------|-------|
+| WER | 10.51 | 10.62 |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+| lm_scale |
+|----------|
+|   0.2    |
+
+To reproduce the above result, use the training commands above, and the following command for decoding:
+
+```
+./conformer_ctc/decode.py \
+  --epoch 18 \
+  --avg 6 \
+  --method whole-lattice-rescoring \
+  --num-paths 1000 \
+  --exp-dir conformer_ctc/exp_500 \
+  --lang-dir data/lang_bpe_500 \
+  --max-duration 20 \
+  --num-workers 1
+```
+Note: the `whole-lattice-rescoring` method is about twice as fast as the `attention-decoder` method, with slightly worse WER.
+
+Pretrained model is available at
+<https://huggingface.co/wgb14/icefall-asr-gigaspeech-conformer-ctc>
+
+The tensorboard log for training is available at
+<https://tensorboard.dev/experiment/rz63cmJXSK2fV9GceJtZXQ/>