This commit is contained in:
Dongji Gao 2023-09-19 14:23:27 -04:00
commit 8c5c23ebe7

View File

@ -1,7 +1,7 @@
# Introduction # Introduction
This is a weakly supervised ASR recipe for the LibriSpeech (clean 100 hours) dataset. We train a This is a weakly supervised ASR recipe for the LibriSpeech (clean 100 hours) dataset. We train a
conformer model using Bypass Temporal Classification (BTC)/Omni-temporal Classification (OTC) with transcripts with synthetic errors. In this README, we will describe conformer model using [Bypass Temporal Classification](https://arxiv.org/pdf/2306.01031.pdf) (BTC)/Omni-temporal Classification (OTC) with transcripts with synthetic errors. In this README, we will describe
the task and the BTC/OTC training process. the task and the BTC/OTC training process.
Note that OTC is an extension of BTC and supports all BTC functions. Therefore, in the following, we only describe OTC. Note that OTC is an extension of BTC and supports all BTC functions. Therefore, in the following, we only describe OTC.
@ -124,13 +124,26 @@ The bypass arc deals with substitution and insertion errors, while the self-loop
### Decoding ### Decoding
``` ```
export CUDA_VISIBLE_DEVICES="0" export CUDA_VISIBLE_DEVICES="0"
python conformer_ctc2/decode.py \ ./conformer_ctc2/decode.py \
--exp-dir "${exp_dir}" \ --exp-dir "${exp_dir}" \
--lang-dir "${otc_lang_dir}" \ --lang-dir "${otc_lang_dir}" \
--lm-dir "${lm_dir}" \ --lm-dir "${lm_dir}" \
--otc-token "${otc_token}" --otc-token "${otc_token}"
``` ```
### Results (ctc-greedy-search)
| Traning Criterion | test-clean | test-other |
|------------|:-------:|:----:|
| CTC |100.0|100.0 |
| OTC | 11.89 | 25.46 |
### Results (1best, blank_bias=-4)
| Traning Criterion | test-clean | test-other |
|------------|:-------:|:----:|
| CTC |98.40|98.68 |
| OTC | 6.59 | 15.98 |
## Citations ## Citations
``` ```
@article{gao2023bypass, @article{gao2023bypass,