I'm trying to train some small ctc models and combine their outputs, but I just realized the merging is not very straightforward, arithmetic mean is obviously not a good idea.
\nI can only find someone's post but no satisfying answer.
\nSo is there someone who have any idea? Thanks
", "messages": [ { "id": 2382861, "postDate": "2023-08-10T04:35:41.040Z", "content": "do you need a language model?
\ni am partcipating the bengali speech recognition.
\nThere , we use CTC to predict charcters first.
\nthen we have a language model guided beam search to turn charcters to word.
I'm trying to train some small ctc models and combine their outputs, but I just realized the merging is not very straightforward, arithmetic mean is obviously not a good idea.
\nI can only find someone's post but no satisfying answer.
\nSo is there someone who have any idea? Thanks
", "rawMarkdown": "I'm trying to train some small ctc models and combine their outputs, but I just realized the merging is not very straightforward, arithmetic mean is obviously not a good idea. \n\nI can only find [someone's post](https://stackoverflow.com/questions/55702043/combining-the-outputs-of-multiple-models-with-ctc-output-layerensemble) but no satisfying answer.\n\nSo is there someone who have any idea? Thanks", "votes": 1 }, { "id": 2381478, "postDate": "2023-08-09T08:01:55.290Z", "content": "In this work, element-wise multiplication is used. It seems to me that it is best to start with the first and go to the last option.https://www.kaggle.com/code/aleksandrkruchinin/tflite-ensemble
", "rawMarkdown": "In this work, element-wise multiplication is used. It seems to me that it is best to start with the first and go to the last option.[https://www.kaggle.com/code/aleksandrkruchinin/tflite-ensemble](url)", "votes": 2 }, { "id": 2378505, "postDate": "2023-08-07T17:02:11.120Z", "content": "I have tried combine by simple avg, it improved only a bit.
", "rawMarkdown": "I have tried combine by simple avg, it improved only a bit.", "votes": 2 } ], "comments": [ { "id": 2382861, "author_name": "hengck23", "author_url": "", "post_date": "2023-08-10T04:35:41.040000", "content": "do you need a language model?
\ni am partcipating the bengali speech recognition.
\nThere , we use CTC to predict charcters first.
\nthen we have a language model guided beam search to turn charcters to word.
In this work, element-wise multiplication is used. It seems to me that it is best to start with the first and go to the last option.https://www.kaggle.com/code/aleksandrkruchinin/tflite-ensemble
", "votes": 2, "replies": [] }, { "id": 2378505, "author_name": "gezi", "author_url": "", "post_date": "2023-08-07T17:02:11.120000", "content": "I have tried combine by simple avg, it improved only a bit.
", "votes": 2, "replies": [] } ], "raw_markdown_by_id": { "2382861": "do you need a language model?\ni am partcipating the bengali speech recognition.\nThere , we use CTC to predict charcters first.\nthen we have a language model guided beam search to turn charcters to word.", "2378227": "I'm trying to train some small ctc models and combine their outputs, but I just realized the merging is not very straightforward, arithmetic mean is obviously not a good idea. \n\nI can only find [someone's post](https://stackoverflow.com/questions/55702043/combining-the-outputs-of-multiple-models-with-ctc-output-layerensemble) but no satisfying answer.\n\nSo is there someone who have any idea? Thanks", "2381478": "In this work, element-wise multiplication is used. It seems to me that it is best to start with the first and go to the last option.[https://www.kaggle.com/code/aleksandrkruchinin/tflite-ensemble](url)", "2378505": "I have tried combine by simple avg, it improved only a bit." } }