This repo contains instructions and scripts to train acoustic models using Kaldi over the datasets of the FalaBrasil Group in Brazilian Portuguese.

🦊 Looking for speech datasets in Brazilian Portuguese? Check out our "Audio Corpora" GitLab group: https://gitlab.com/fb-audio-corpora

🦊 Looking for language models or phonetic dictionaries? Check out our "NLP resources" GitLab group: https://gitlab.com/fb-nlp

☕ Looking for Kaldi installation instructions? Check out our install guide on INSTALL.md file or just go follow Kaldi documentation directly: https://github.com/kaldi-asr/kaldi

See fb-mini_librispeech/ dir. Based on Mini-librispeech nnet3 recipe (local/chain/tuning/run_tdnn_1j.sh).

$ ./prep_minilibri.sh /path/to/kaldi/egs/myproject
$ cd /path/to/kaldi/egs/myproject/s5/
$ ./run.sh

For online decoding, please check fb-mini_librispeech/fbvosk/ dir. Dir utils/online/ is deprecated.

See fb-aspire/ dir. Based on ASpIRE nnet3 recipe.

$ ./prep_aspire.sh /path/to/kaldi/egs/myproject
$ cd /path/to/kaldi/egs/myproject/s5/
$ ./run.sh

See fb-librispeech/ dir. Based on LibriSpeech nnet3 recipe.

$ ./prep_libri.sh /path/to/kaldi/egs/myproject
$ cd /path/to/kaldi/egs/myproject/s5/
$ ./run_all.sh

⚠️ These scripts are experimental for forced phonetic alignment. For transcription you may stick with Mini-libri recipe.

See fb-callhome/ dir. Based on CALLHOME v2 recipe.

$ ./prep_callhome.sh /path/to/kaldi/egs/myproject
$ cd /path/to/kaldi/egs/myproject/v2/
$ ./run.sh

Standalone clustering procedure based on pyanote-audio lib can also be found under utils/clustering/ dir.

If you use these codes or want to mention the paper referred above, please cite us as one of the following:

Batista, C., Dias, A.L., Sampaio Neto, N. (2018) Baseline Acoustic Models for Brazilian Portuguese Using Kaldi Tools. Proc. IberSPEECH 2018, 77-81, DOI: 10.21437/IberSPEECH.2018-17.

@inproceedings{Batista2018,
  author    = {Cassio Batista and Ana Larissa Dias and Nelson {Sampaio Neto}},
  title     = {{Baseline Acoustic Models for Brazilian Portuguese Using Kaldi Tools}},
  year      = {2018},
  booktitle = {Proc. IberSPEECH 2018},
  pages     = {77--81},
  doi       = {10.21437/IberSPEECH.2018-17},
  url       = {http://dx.doi.org/10.21437/IberSPEECH.2018-17}
}

⚠️ This paper uses the outdated nnet2 recipes, while this repo has been updated to the chain models' recipe via nnet3 scripts. If you really want nnet2 scripts, you may find them on tag nnet2. Try running git tag.

Dias A.L., Batista C., Santana D., Neto N. (2020) Towards a Free, Forced Phonetic Aligner for Brazilian Portuguese Using Kaldi Tools. In: Cerri R., Prati R.C. (eds) Intelligent Systems. BRACIS 2020. Lecture Notes in Computer Science, vol 12319. Springer, Cham. https://doi.org/10.1007/978-3-030-61377-8_44

@inproceedings{Dias20,
  author     = {Dias, Ana Larissa and Batista, Cassio and Santana, Daniel and Neto, Nelson},
  editor     = {Cerri, Ricardo and Prati, Ronaldo C.},
  title      = {Towards a Free, Forced Phonetic Aligner for Brazilian Portuguese Using Kaldi Tools},
  booktitle  = {Intelligent Systems},
  year       = {2020},
  publisher  = {Springer International Publishing},
  address    = {Cham},
  pages      = {621--635},
  isbn       = {978-3-030-61377-8}
}

Coming soon.

Grupo FalaBrasil (2021) - https://ufpafalabrasil.gitlab.io/
Universidade Federal do Pará (UFPA) - https://portal.ufpa.br/
Cassio Batista - https://cassota.gitlab.io/