2022-05-31

A new release with better error rates, largely from the same data as the previous one.

Metrics*:

1. Raw acoustic model (without a scorer)

• Czech Commonvoice 6.1 test dataset: WER: 0.405500, CER: 0.106870, loss: 15.227368
• Vystadial 2016 test dataset: WER: 0.506131, CER: 0.195149, loss: 17.695986
• Large Corpus of Czech Parliament Plenary Hearings test dataset: WER: 0.213377, CER: 0.052676, loss: 20.449242
• ParCzech 3.0 test dataset: WER: 0.209651, CER: 0.061622, loss: 28.217770

2. With the attached czech-large-vocab.scorer:

• Czech Commonvoice 6.1 test dataset: WER: 0.152865, CER: 0.067557, loss: 15.227368**
• Vystadial 2016 test dataset: WER: 0.357435, CER: 0.201479, loss: 17.695986
• Large Corpus of Czech Parliament Plenary Hearings test dataset: WER: 0.097380, CER: 0.036706, loss: 20.449242
• ParCzech 3.0 test dataset: WER: 0.101289, CER: 0.045102, loss: 28.217770

Metrics for the quantized model are circa one percent worse.

*Any clips longer than thirty seconds were discarded
**Better than expected results on the common voice set with the language model might possibly be explained by a partial overlap of the test transcriptions and language model sources, namely Wikipedia and Europarl v7.

2021-07-21

This is a model based on a smaller alphabet that only contains Czech alphabet letters (as opposed to noise and non-speech sound symbols), see the file alphabet.txt.

Results on some test sets (without a language model):

• Czech Commonvoice 6.1 test dataset: WER: 0.423823, CER: 0.112101, loss: 15.059019
• Vystadial 2016 test set: WER: 0.507822, CER: 0.195558, loss: 17.671772
• Large Corpus of Czech Parliament Plenary Hearings test set: WER: 0.214612, CER: 0.051837, loss: 19.688087

2021-04-08

Results on some test sets (without a language model):

• Czech Commonvoice test dataset: WER: 0.446331, CER: 0.112392, loss: 14.529331
• Vystadial 2016 test set: WER: 0.569942, CER: 0.226371, loss: 20.126215
• Large Corpus of Czech Parliament Plenary Hearings test set: WER 0.209104, CER: 0.048405, loss: 17.649645

While not quite SOTA, it may be a sufficient basis for recognition with limited vocabulary.