Extending Allosaurus for Australian languages

intro.md

Allosaurus

• Allosaurus is a pretrained universal phone recognizer: https://github.com/xinjli/allosaurus
• It has been trained on English, Japanese, Mandarin, Tagalog, Turkish, Vietnamese, German, Spanish, Amharic, Italian and Russian

Testing on Kaytetye

• We test this off-the-shelf version on some Kaytetye data. The data are citation form headwords recorded by a female native speaker of Kaytetye in a music studio for a multimedia dictionary.
• For 2,360 headwords, we have ~2 repetitions per word (e.g. palpalpe), and two transcriptions (t_1, t_2) by two independent human transcribers.
• In the table below int_t_dist is the inter-transcriber string distance and min_a_dist is the minimum string distance between the allosaurus transcription and the human transcriptions.

word	rep	t_1	t_2	allosaurus	int_t_dist	min_a_dist
kwake	2	kwakə	kwakə	kwakə	0	0
etelepwenke	1	itələpunk	itələpunk	itələpunk	0	0
itanperre	1	itanpəɾə	itanpəɾə	ətaɳpəɾə	0	2
palpalpe	2	paɭpaɭpə	palpalp	palʲpal	3	2
mwanyeme	1	mutɲəmə	muɲim	mwan	3	4
mwanyeme	2	muɲəmə	muɲim	mwanam	2	4
ngkwerengke	1	ŋkuɺəŋk	ŋkwuɺəŋk	pulan	1	6
ngkwerengke	2	ŋkuɺəŋk	ŋkwuɺəŋk	puɻaŋk	1	4

• The figure below shows the comparison between int_t_dist (x-axis) and min_a_dist (y-axis) for 4741 tokens (2360 types). Overall, human transcribers generally agree with each other (most human transcriptions are 0 or 1).
• The median allosaurus transcription is at least 4 characters off from the human transcriptions.

Extending Allosaurus

• One way to extend Allosaurus for Australian languages might be to include a Dravidian language in the training data.
• Many have noted superficial similarities in the phonological systems between Dravidian languages of South India (e.g. Tamil) and the Pama-Nyungan languages of Australia (e.g. Kaytetye).
• 5-6 places of articulation, for both stops [p t̪ t t (c) k] and nasals [m n̪ n ɳ (ɲ) ŋ]
• No voicing contrast [p, b] ∈ /p/

Tamil dataset

• He et al. (2020) have recently released a "Crowdsourced high-quality Tamil multi-speaker speech data set": http://www.openslr.org/65/ with 50 speakers (25F) and ~8 hours of audio (~4F)

• Each .wav file has an acommpanying transcript in Tamil script, e.g. for tag_09162_01279273055.wav:

  • அவர்களின் படங்களின் டீஸருக்கு கிடைக்கும் வரவேற்பு அபிரிதமாக உள்ளது

• Epitran, the grapheme-to-phoneme conversion tool, already has a Tamil script conversion function:

  >>> import epitran
  >>> epi = epitran.Epitran('tam-Taml')
  >>> print(epi.transliterate(u'அவர்களின் படங்களின் டீஸருக்கு கிடைக்கும் வரவேற்பு அபிரிதமாக உள்ளது'))
  aʋaɾkaɭin paʈaŋkaɭin ʈiːsaɾukku kiʈajkkum ʋaɾaʋeːrpu apiɾit̪amaːka uɭɭat̪u
  

• Allovera https://github.com/dmort27/allovera the tool for phoneme-to-allophone mappings used by Allosaurus does not yet have a Tamil mapping, though the JSON format looks relatively straightforward:

{
    "iso": "eng",
    "glottocodes": ["nort3312"],
    "primary src": "Ladefoged:1999-american",
    "secondary srcs": [],
    "epitran": "eng-Latn",
    "mappings": [
        {
            "phone": "pʰ",
            "phoneme": "p",
            "environment": "word initially; in onset of stressed syllables"
        },
        {
            "phone": "p",
            "phoneme": "p",
            "environment": "elsewhere"
        },
...

• So the current task is to create a draft phoneme-to-allophone mapping from a reference grammar and check it with a linguist who is familiar with Tamil phonology.

• At the same time, I will need to find out how to train Allosaurus (the Git repo only seems to provide the pre-trained model).