search

alphacep / vosk-api

Offline speech recognition API for Android, iOS, Raspberry Pi and servers with Python, Java, C# and Node
Apache License 2.0
7.7k stars 1.08k forks source link

HCLr.fst and final.mdl decode a wav file dynamicly is none !!! #1293

Open wwbnjsace opened 1 year ago

wwbnjsace commented 1 year ago

i have two question ,the first : when i use "vosk-api-master/training/local/chain/run_tdnn.sh" train a model ,the model size is about 20M but the release model size is 70M (vosk-model-en-us-0.22-lgraph ,can u tell me why?

the second : when i use my 20M model ,i decode a file in final.mdl and HCLG.fst ,the output is normal ; but when i utils/mkgraph_lookahead.sh generate the HCLr.fst ,i decode a file by final and HCLr.st in a dynamicly way by language model,the output is none , can you tell me why?

nshmyrev commented 1 year ago

when i use "vosk-api-master/training/local/chain/run_tdnn.sh" train a model ,the model size is about 20M but the release model size is 70M (vosk-model-en-us-0.22-lgraph ,can u tell me why?

Training creates a toy model, you need much more data to train a serious model

i decode a file by final and HCLr.st in a dynamicly way by language model,the output is none , can you tell me why?

Something mismatch. There could be many reasons, like wrong words.txt which should be different for HCLG.

wwbnjsace commented 1 year ago

when i use "vosk-api-master/training/local/chain/run_tdnn.sh" train a model ,the model size is about 20M but the release model size is 70M (vosk-model-en-us-0.22-lgraph ,can u tell me why?

Training creates a toy model, you need much more data to train a serious model

i decode a file by final and HCLr.st in a dynamicly way by language model,the output is none , can you tell me why?

Something mismatch. There could be many reasons, like wrong words.txt which should be different for HCLG.

i use about 7000 hours data to train the model ,and the model size has no relationship with the training data，is it ? AND the Something mismatch i find there is no mismatch ;when i only use the HCLG.fst replace the HCLr.fst ,the output is normal .

nshmyrev commented 1 year ago

i use about 7000 hours data to train the model ,and the model size has no relationship with the training data，is it ?

What language are you training for?

when i only use the HCLG.fst replace the HCLr.fst ,the output is normal .

It exactly means you didn't pick words.txt. Make sure you remove words.txt or use words.txt from the lgraph. The one in graph is different.

wwbnjsace commented 1 year ago

i use about 7000 hours data to train the model ,and the model size has no relationship with the training data，is it ?

What language are you training for?

when i only use the HCLG.fst replace the HCLr.fst ,the output is normal .

It exactly means you didn't pick words.txt. Make sure you remove words.txt or use words.txt from the lgraph. The one in graph is different.

Chinese English ，without ivector . And when i add ivector ,the model size is became a little big ,about 25M .