From ee2fa42e5d537d4aab3725ab78a6224dcecd7a1e Mon Sep 17 00:00:00 2001 From: marcoyang1998 Date: Mon, 31 Jul 2023 14:18:58 +0800 Subject: [PATCH] minor fix --- .../decoding-with-langugage-models/LODR.rst | 50 +++++++++---------- 1 file changed, 25 insertions(+), 25 deletions(-) diff --git a/docs/source/decoding-with-langugage-models/LODR.rst b/docs/source/decoding-with-langugage-models/LODR.rst index e0ca13a99..b6625ee1d 100644 --- a/docs/source/decoding-with-langugage-models/LODR.rst +++ b/docs/source/decoding-with-langugage-models/LODR.rst @@ -4,59 +4,59 @@ LODR for RNN Transducer ======================= -As a type of E2E model, neural transducers are usually considered as having an internal -language model, which learns the language level information on the training corpus. -In real-life scenario, there is often a mismatch between the training corpus and the target corpus space. +As a type of E2E model, neural transducers are usually considered as having an internal +language model, which learns the language level information on the training corpus. +In real-life scenario, there is often a mismatch between the training corpus and the target corpus space. This mismatch can be a problem when decoding for neural transducer models with language models as its internal language can act "against" the external LM. In this tutorial, we show how to use `Low-order Density Ratio `_ to alleviate this effect to further improve the performance -of langugae model integration. +of langugae model integration. .. note:: - This tutorial is based on the recipe + This tutorial is based on the recipe `pruned_transducer_stateless7_streaming `_, - which is a streaming transducer model trained on `LibriSpeech`_. + which is a streaming transducer model trained on `LibriSpeech`_. However, you can easily apply LODR to other recipes. If you encounter any problems, please open an issue here `icefall `__. .. note:: - For simplicity, the training and testing corpus in this tutorial are the same (`LibriSpeech`_). However, - you can change the testing set to any other domains (e.g `GigaSpeech`_) and prepare the language models + For simplicity, the training and testing corpus in this tutorial are the same (`LibriSpeech`_). However, + you can change the testing set to any other domains (e.g `GigaSpeech`_) and prepare the language models using that corpus. -First, let's have a look at some background information. As the predecessor of LODR, Density Ratio (DR) is first proposed `here `_ +First, let's have a look at some background information. As the predecessor of LODR, Density Ratio (DR) is first proposed `here `_ to address the language information mismatch between the training corpus (source domain) and the testing corpus (target domain). Assuming that the source domain and the test domain are acoustically similar, DR derives the following formular for decoding with Bayes' theorem: .. math:: - \text{score}\left(y_u|\mathit{x},y\right) = - \log p\left(y_u|\mathit{x},y_{1:u-1}\right) + - \lambda_1 \log p_{\text{Target LM}}\left(y_u|\mathit{x},y_{1:u-1}\right) - + \text{score}\left(y_u|\mathit{x},y\right) = + \log p\left(y_u|\mathit{x},y_{1:u-1}\right) + + \lambda_1 \log p_{\text{Target LM}}\left(y_u|\mathit{x},y_{1:u-1}\right) - \lambda_2 \log p_{\text{Source LM}}\left(y_u|\mathit{x},y_{1:u-1}\right) -where :math:`\lambda_1` and :math:`\lambda_2` are the weights of LM scores for target domain and source domain respectively. -Here, the source domain LM is trained on the training corpus. The only difference in the above formular compared to +where :math:`\lambda_1` and :math:`\lambda_2` are the weights of LM scores for target domain and source domain respectively. +Here, the source domain LM is trained on the training corpus. The only difference in the above formular compared to shallow fusion is the subtraction of the source domain LM. -Some works treat the predictor and the joiner of the neural transducer as its internal LM. However, the LM is +Some works treat the predictor and the joiner of the neural transducer as its internal LM. However, the LM is considered to be weak and can only capture low-level language information. Therefore, `LODR `__ proposed to use a low-order n-gram LM as an approximation of the ILM of the neural transducer. This leads to the following formula during decoding for transducer model: .. math:: - \text{score}\left(y_u|\mathit{x},y\right) = - \log p_{rnnt}\left(y_u|\mathit{x},y_{1:u-1}\right) + - \lambda_1 \log p_{\text{Target LM}}\left(y_u|\mathit{x},y_{1:u-1}\right) - + \text{score}\left(y_u|\mathit{x},y\right) = + \log p_{rnnt}\left(y_u|\mathit{x},y_{1:u-1}\right) + + \lambda_1 \log p_{\text{Target LM}}\left(y_u|\mathit{x},y_{1:u-1}\right) - \lambda_2 \log p_{\text{bi-gram}}\left(y_u|\mathit{x},y_{1:u-1}\right) -In LODR, an additional bi-gram LM estimated on the source domain (e.g training corpus) is required. Comared to DR, +In LODR, an additional bi-gram LM estimated on the source domain (e.g training corpus) is required. Comared to DR, the only difference lies in the choice of source domain LM. According to the original `paper `_, LODR achieves similar performance compared DR in both intra-domain and cross-domain settings. As a bi-gram is much faster to evaluate, LODR is usually much faster. @@ -99,17 +99,17 @@ The following WERs are achieved on test-clean and test-other: $ For test-other, WER of different settings are: $ beam_size_4 7.93 best for test-other -Then, we download the external language model and bi-gram LM that are necessary for LODR. +Then, we download the external language model and bi-gram LM that are necessary for LODR. Note that the bi-gram is estimated on the LibriSpeech 960 hours' text. .. code-block:: bash $ # download the external LM - $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm + $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm $ # create a symbolic link so that the checkpoint can be loaded $ pushd icefall-librispeech-rnn-lm/exp $ git lfs pull --include "pretrained.pt" - $ ln -s pretrained.pt epoch-99.pt + $ ln -s pretrained.pt epoch-99.pt $ popd $ $ # download the bi-gram @@ -122,7 +122,7 @@ Note that the bi-gram is estimated on the LibriSpeech 960 hours' text. Then, we perform LODR decoding by setting ``--decoding-method`` to ``modified_beam_search_lm_LODR``: .. code-block:: bash - + $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp $ lm_dir=./icefall-librispeech-rnn-lm/exp $ lm_scale=0.42 @@ -135,7 +135,7 @@ Then, we perform LODR decoding by setting ``--decoding-method`` to ``modified_be --exp-dir $exp_dir \ --max-duration 600 \ --decode-chunk-len 32 \ - --decoding-method modified_beam_search_lm_LODR \ + --decoding-method modified_beam_search_LODR \ --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model \ --use-shallow-fusion 1 \ --lm-type rnn \ @@ -181,4 +181,4 @@ indeed **further improves** the WER. We can do even better if we increase ``--be - 6.38 * - 12 - 2.4 - - 6.23 \ No newline at end of file + - 6.23