rasr release: add decoding results

Summary:
- add links
- add decoding results (except swb for now)
- add other ddecoder params

Reviewed By: xuqiantong

Differential Revision: D25700435

fbshipit-source-id: 0ebd104eae8c4bd1d773e04dfaf9a6ab46901ad7

recipes/rasr/README.md

@@ -5,13 +5,14 @@ This is a repository sharing pre-trained acoustic models and language models for
 
 ## Dependencies
 
-* [flashlight](https://github.com/facebookresearch/flashlight)
+* [`Flashlight`](https://github.com/facebookresearch/flashlight)
+* [`Flashlight` ASR app](https://github.com/facebookresearch/flashlight/tree/master/flashlight/app/asr)
 
 ## Models
 
 ### Acoustic Model
 
-All the acoustic models are retrained using flashlight with [wav2letter++](https://github.com/facebookresearch/wav2letter) consolidated. `Tedlium` is not used as training data here due to license issue. All the training data has more standardized sample rate 16kHz rather than 8kHz used in the paper.
+All the acoustic models are retrained using `Flashlight` with [wav2letter++](https://github.com/facebookresearch/wav2letter) consolidated. `Tedlium` is not used as training data here due to license issue. All the training data has more standardized sample rate 16kHz rather than 8kHz used in the paper.
 
 Here, we are releasing models with different architecture and different sizes. Note that the models may not fully reproduce results in the paper because of both data and toolkit implementation discrepancies.
 
@@ -44,32 +45,32 @@ The perplexities of the LMs on different development sets are listed below.
 
 ### WER
 
-Here we summarize the decoding WER for all releasing models. All the numbers in the table are in format `viterbi WER -> beam search WER`.
+Here we summarize the decoding WER for all releasing models. All the numbers in the table are in format `viterbi WER -> beam search WER (small beam/large beam)`.
 
 |Achitecture |# Param |nov92 |TL-test |CV-test |LS-test-clean |LS-test-other |Hub05-SWB |Hub05-CH |
 | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
-|Transformer |300 mil |3.4 → 2.9 |7.6 → 5.5 |15.5 → 11.6 |3.0 → 3.2 |7.2 → 6.4 |6.8 |11.6 |
-|Transformer |70 mil |4.5 |9.4 |19.8 |4 |9.7 |7.5 |13 |
-|Conformer |300 mil |3.5 |8.4 |17 |3.2 |8 |7 |11.9 |
-|Conformer |87 mil |4.3 |8.7 |18.2 |3.7 |8.6 |7.3 |12.2 |
-|Conformer |28 mil |5 |10.5 |22.2 |4.7 |11.1 |8.8 |13.7 |
+|Transformer |300 mil |3.4 → 2.9/2.9 |7.6 → 5.5/5.4 |15.5 → 11.6/11.2 |3.0 → 3.2/3.2 |7.2 → 6.4/6.4 |6.8 → 6.2/6.2 |11.6 → 10.8/10.7 |
+|Transformer |70 mil |4.5 → 3.7/3.5 |9.4 → 6.2/6.1 |19.8 →13.8/13.0 |4 → 3.6 /3.6 |9.7 → 7.7/7.5 |7.5 → 6.6/6.5 |13 → 11.8/11.7 |
+|Conformer |300 mil |3.5 → 3.3/3.3 |8.4 → 6.2/6.0 |17 → 12.7/12.0 |3.2 → 3.4/3.4 |8 → 7/6.8 |7 → 6.4/6.5 |11.9 → 10.7/10.5 |
+|Conformer |87 mil |4.3 → 3.3/3.3 |8.7 → 6.1/5.9 |18.2 →13.1/12.4 |3.7 → 3.5/3.5 |8.6 → 7.4/7.2 |7.3 → 6.7/6.7 |12.2 → 11.7/11.5 |
+|Conformer |28 mil |5 → 3.9/3.8 |10.5 → 6.9/6.6 |22.2 → 15.4/14.4 |4.7 → 4/3.9 |11.1 → 8.9/8.6 |8.8 → 7.8/7.7 |13.7 → 12.4/12.2 |
 
 Decoding is done with lexicon-based beam-search decoder using 200k common crawl lexicon and small common crawl lm.
 * [tokens](https://[dl.fbaipublicfiles.com/wav2letter/rasr/tutorial/tokens.txt](http://dl.fbaipublicfiles.com/wav2letter/rasr/tutorial/tokens.txt))
 * [inference lexicon](https://dl.fbaipublicfiles.com/wav2letter/rasr/tutorial/lexicon.txt)
-* Decoding parameters:
+* Decoding parameters (`beamthreshold=100, beamsizetoken=30`):
 
 |Achitecture |# Param |LM Weight |Word Score |Beam Size |
 | :---: | :---: | :---: | :---: | :---: |
-|Transformer |300 mil |1.5 |0 |50 |
-|Transformer |70 mil | | | |
-|Conformer |300 mil | | | |
-|Conformer |87 mil | | | |
-|Conformer |28 mil |2 |0 |50 |
+|Transformer |300 mil |1.5 |0 |50/500 |
+|Transformer |70 mil |1.7 |0 |50/500 |
+|Conformer |300 mil |1.8 |2 |50/500 |
+|Conformer |87 mil |2 |0 |50/500 |
+|Conformer |28 mil |2 |0 |50/500 |
 
 ## Tutorial
 
-To simply serialize all the model and interact with them, please refer to the Flashlight tutorials as in [here](https://github.com/facebookresearch/flashlight/tree/master/flashlight/app/asr/tutorial).
+To simply serialize all the models and interact with them, please refer to the [`Flashlight` ASR app tutorials](https://github.com/facebookresearch/flashlight/tree/master/flashlight/app/asr/tutorial).