As a final project for a course on Speech Technology, I used the Montreal Forced Aligner v1.0.1 to train an acoustic model for Catalan. I’m sharing this, without any guarantee of any kind, in the hopes that it may help those that do phonetic/phonological research on Catalan. The model was trained on the FestCat corpus. You can download the trained acoustic model by clicking here.
The G2P model produces some errors, some of which are outlined below. If you are a researcher with access to an improved G2P model for Catalan, it would be great if you reached out to me so we can reproduce the automated transcription of the corpus and train an improved acoustic model.
G2P
The G2P model I used is based on espeak. This means it produces several symbols that I have not seen used before when phonetically transcribing Catalan, as well as other quirks. After generating a pronunciation dictionary for your data, I recommend manually correcting the transcriptions. Some things that I noticed initially:
- <rr> produces an alveolar tap followed by a trill instead of just a trill, i.e. [ɾ r].
- [ɐ] instead of [ə]
- [ʊ] instead of [u]
- [ɕ] instead of [ʃ]
- [ʑ] instead of [ʒ]
Other errors in the G2P model passed along to me by Annie Helms:
- word-final ‘-ent’ as ‘e n t’ versus ‘e n’ (dialect dependent)
- word-final ‘-or’ as ‘o r’ versus ‘o’ (dialect dependent)
- word-final ‘-lt’ as ‘l t’ instead of ‘l’ (dialect dependent)
- ‘ix’ as ‘j ʃ’ instead of just ‘ ʃ’
- ‘g i’ sequence as ‘ɣ i’ instead of ‘ʒ/ʑ’
- ‘v’ as ‘v’ instead of ‘b/β’ (dialect dependent)
- spirantization after vowels (i.e., from ‘e g z’ to ‘e ɣ z’) (dialect dependent)
- ‘aquest/aquests’ as ‘a k ɛ s t (s)’ instead of ‘a k ɛ t (s)’ (dialect dependent)
- ‘qui’ as ‘k w i’ instead of ‘k i’
- ‘ningú’ as ‘n u’ instead of ‘n g u’ (dialect dependent)
- ‘extens’ as ‘ɛ x t ə n s’ instead of ‘ə x t ɛ n s’
- ‘complagut’ as ‘m l’ instead of ‘m p l’