rhasspy / en-us_kaldi-rhasspy

Kaldi profile for U.S. English trained from open speech data

U.S. English Kaldi Profile

A Rhasspy profile for U.S. English (en).

Includes:

• A Kaldi nnet3 speech to text model
  • See files in acoustic_model/
  • Recipe created with ipa2kaldi
  • Word error rate: 5.90%
  • Trained on:
    • Common Voice (1,449 hours)
    • LibriSpeech (363 hours)
    • M-AILabs (98 hours)
    • Tatoeba (185 hours)
    • VCTK (41 hours)
    • Voxforge (124 hours)
• An IPA pronunciation lexicon based on Reece Dunn's U.S. English lexicon
  • See base_dictionary.txt.gz
• A phonetisaurus grapheme to phoneme model for predicting word pronunciations
  • See g2p.fst.gz
  • Trained on base_dictionary.txt.gz
• A tri-gram ARPA language model
  • See base_language_model.txt.gz
  • Text from audio transcriptions, Common Voice, and Universal Dependencies