- Scorer KenLM Scorer in Esperanto
- Deepspeech/Coqui AI models - find and Download .tflite models
- Colab Notebooks
Using deepspeech/coqui ai and the common voice dataset
- https://github.com/coqui-ai/STT
- https://stt.readthedocs.io/en/latest/
- Tim jam kreis sistemon en esperantto: https://gitlab.com/54696d21
Tools/Iloj
- Ilo por krei subtekstojn en Esperanto
Video: Kiel krei aŭtomatajn subtekstojn en Esperanto per Google Colab
| Datumaro | versio | grandeco | permesilo |
|---|---|---|---|
| Common Voice | CV Corpus 7.0 | 17 GB 748 h | CC 0 |
| tatoeba | 03.06.20 | 4 063 audio files | CC-BY |
| lingualibre | 03.06.20 | 425 MB | CC BY-SA |
| datumaro | parametroj | GPU | rezultoj |
|---|---|---|---|
| eo_41h_2019-12-10 | ? | 2 x 1080 Ti 32Gb RAM (leadertelecom) | WER 0.5 |
| eo_844h_2021-07-21 | english checkpoints, n_depth 2048, dropout_rate 0.3, learning_rate 0.0001 details | Google Colab Pro Plus | WER 24,7% (test was part of train dataset) download |
There is an Esperanto Vosk Model that can be used in many tools such as Kdenlive to create subtitles: https://alphacephei.com/vosk/models
- run ssh process in background next time (background + disown process)
- experiment with different data tables e.g. ignore sentences with one no-vote
- extract the Tatoeba corpus with the script from https://github.com/DanBmh/deepspeech-german
- Extract lingua libre files https://github.com/mozilla/DeepSpeech/blob/master/bin/import_lingua_libre.py
- create kenlm language model (scorer). https://tiefenauer.github.io/blog/wiki-n-gram-lm/ https://github.com/kpu/kenlm