donaldos
commented
6 days ago Directory structure
. ├── am │ ├── final.mdl │ └── tree ├── conf │ ├── mfcc.conf │ └── model.conf ├── graph │ ├── Gr.fst │ ├── HCLr.fst │ ├── disambig_tid.int │ ├── phones │ │ ├── align_lexicon.int │ │ ├── align_lexicon.txt │ │ ├── disambig.int │ │ ├── disambig.txt │ │ ├── optional_silence.csl │ │ ├── optional_silence.int │ │ ├── optional_silence.txt │ │ ├── silence.csl │ │ ├── word_boundary.int │ │ └── word_boundary.txt │ ├── phones.txt │ └── words.txt └── ivector ├── final.dubm ├── final.ie ├── final.mat ├── global_cmvn.stats ├── online_cmvn.conf └── splice.conf
nshmyrev
Dear Nickolay V. Shmyrev
I have been using the vosk api to generate and recognise dynamic grammars well. In particular, I have been testing a lot with a customised engine configuration by generating HCLr.fst and Gr.fst using compile-graph.sh. The model I used was based on the final.mdl file of ‘vosk-model-small-en-us-0.15’ and generated the final file. For the acoustic model upgrade, we used the final.mdl of vosk-model-en-us-0.22 to generate HCLr.fst and Gr.fst.
INFO (2024-11-20 09:25:06,800:main): [EnumaeduSREngineFile.py:57] - proc_sr() - [‘sit down please’, ‘[unk]’] LOG (VoskAPI:UpdateGrammarFst():recogniser.cc:287) [‘sit down please’, ‘[unk]’] LOG (VoskAPI:Estimate():language_model.cc:142) Estimating language model with ngram-order=2, discount=0.5 LOG (VoskAPI:OutputToFst():language_model.cc:209) Created language model with 5 states and 10 arcs. As you can see in the log, when you utter the phrase ‘sit down please’, it recognises only [{‘start’: 0.570133, ‘end’: 1.409742, ‘word’: ‘down’, ‘conf’: 1.0}]. However, if you use the full resource vosk-model-en-us-0.22-lgraph, which is the example, you will not encounter any problems.
What could be the cause of this, and is there a methodology to validate it?
and Where do I need to copy resources from? "exp>tdnn>lgraph" or "exp>tdnn>lgraph_orig"