1215thebqtic

fix bug: context score does not affect token's tot_cost

4

solve the issue discussed here #956

context biasing can't work with LatticeFasterOnlineDecoder::GetBestPath()

1

In order for quick online decoding, our ASR engine calls `LatticeFasterOnlineDecoder::GetBestPath()`, not generating lattice, to get temporary transcripts every 1 second. I'm using ctc wfst beam search with context biasing....

微调的数据格式支持segments吗

1

您好， 微调时的数据格式是text和wav.scp，我这里有的数据还有segments文件，即为音频其中的某一段，全部切成小文件有点浪费空间。所以数据格式可以支持segments吗？谢谢！

知识型纠错有什么好的方向

1

你好， 比如说我有一个知识库，比如说“A说了XXX”。我现在写了“B说了XXX”，但根据知识库是A说的，所以要把B改成A。 这种知识型的纠错，有什么比较成熟的解决方案吗？谢谢！

请问基于fairseq的seq2seq的工作还在继续跟进吗

2

你好， 看TODO list上说有在基于fairseq重写seq2seq，请问这部分工作还会继续跟进吗？ 我们使用当前版本的seq2seq训练的模型误纠有点多，和seq2edit的precision相比差好多，recall差不多在同一水平，不过两者融合后可以降一些误纠，请问还有其他的思路在保证TP不降很多的情况下降低误纠吗？谢谢~ 祝工作顺利~

initial decoder input in onnx decoding results in deletion errors

1

Hi, I'm using python scripts to decode onnx models, and I found deletion errors in some testsets (cer = 8.8), especially for the first few words in a sentence. However,...

max_duration for zipformer with about 6800 tokens

15

Hi, I'm doing 80k hours training using the default zipformer of wenetspeech recipe. The token size is about 6800 (Chinese char + English bpe) and GPU memory is 32G. The...

Hi, I'm confused about the pad_length in set_features. Since the frames that are less than chunk_size\*2+7+2\*3 have already been padded in streaming_decode.py, why are we still adding 7+2\*3 when initially...

Are there any code that support multi-task learning?

2

Hi, I'm working on a multi-task learning model. I want to first use speech translation to train a branch, and then use speech recognition data to train the other branch....