From f91cdef8e44e1d0b94f370e5466bf4ed7f4e552f Mon Sep 17 00:00:00 2001 From: dujing Date: Mon, 24 Jul 2023 17:24:41 +0800 Subject: [PATCH] Add realtime CTC-KWS demo in README. --- examples/hi_xiaowen/s0/README.md | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/examples/hi_xiaowen/s0/README.md b/examples/hi_xiaowen/s0/README.md index 964b71e..aad6cdc 100644 --- a/examples/hi_xiaowen/s0/README.md +++ b/examples/hi_xiaowen/s0/README.md @@ -59,4 +59,9 @@ The actual FRR will be lower than the DET curve gives in a given threshold. On some small data KWS tasks, we believe the FSMN-CTC model is more robust compared with the classification model using CE/Max-pooling loss. -For more infomation and results of FSMN-CTC KWS model, you can click [modelscope](https://modelscope.cn/models/damo/speech_charctc_kws_phone-wenwen/summary). \ No newline at end of file +For more infomation and results of FSMN-CTC KWS model, you can click [modelscope](https://modelscope.cn/models/damo/speech_charctc_kws_phone-wenwen/summary). + +For realtime CTC-KWS, we should process wave input on streaming-fashion, +include feature extraction, keyword decoding and detection and some postprocessing. +Here is a [demo](https://modelscope.cn/studios/thuduj12/KWS_Nihao_Xiaojing/summary) in python, +the core code is in wekws/bin/stream_kws_ctc.py, you can refer it to implement the runtime code.