On what kind of datasets does the model trained on?

daanzu / kaldi-active-grammar

Python Kaldi speech recognition with grammars that can be set active/inactive dynamically at decode-time

GNU Affero General Public License v3.0

332 stars 49 forks source link

On what kind of datasets does the model trained on? #38

Open sraghuram90 opened 3 years ago

sraghuram90 commented 3 years ago

What are the datasets does this kaldi active grammar model trained on? If you would have included public datasets, could you name them? The pretrained model which you mentioned, is that Zamia speech model?

JohnDoe02 commented 3 years ago

I was also curious about this. According to here (cf., stage 2) it should be: Librispeech, TEDLIUM, Mozilla's Commonvoice, Tatoeba, Tensorflow's speech_commands.

daanzu commented 3 years ago

Actually, daanzu_multi_en is a partial and unfinished training pipeline. I have ended up working with a heavily modified version of the Zamia pipeline. The datasets are:

Common Voice
Common Voice single word
Librispeech
LJ Speech
M-AILabs
Google Speech Commands
Tatoeba
TedLIUM3
Voxforge
A collection of TTS I generated

zhouyong64 commented 2 years ago

Actually, daanzu_multi_en is a partial and unfinished training pipeline. I have ended up working with a heavily modified version of the Zamia pipeline. The datasets are:

Common Voice

Common Voice single word

Librispeech

LJ Speech

M-AILabs

Google Speech Commands

Tatoeba

TedLIUM3

Voxforge

A collection of TTS I generated

How about kaldi_model_daanzu_20211030-biglm? Also trained on these datasets?

daanzu commented 2 years ago

@zhouyong64

How about kaldi_model_daanzu_20211030-biglm? Also trained on these datasets?

Yes, the new model is trained on the same datasets. The major change is that it now includes models necessary for running g2p_en for local pronunciation generation.