📄 Error analysis of the pronunciation of English consonants

Summary

The paper describes how Faroese speakers mispronounce English consonants and suggests remedies, using British RP (also called SSBE or NRP) as the target accent. No textbooks or descriptions existed for this learner group. The data come from six speakers of the Eysturoyarmál dialect in southern Eysturoy. Each read a passage adapted from Roald Dahl's Charlie and the Chocolate Factory. One key informant also gave a free-speech sample, which showed that reading and spontaneous speech differed very little.

The authors sorted 3,547 sound tokens into a "native-like" category or into error categories specific to each consonant. They group the errors into three kinds:

• Phonemic: for example /θ/→/t/ and /w/→/v/. Many non-native speakers worldwide make these.
• Allophonic: these are unusual and typically Faroese, such as pre-aspirated fortis stops and devoiced sonorants.
• Distributional: mainly rhoticity.

Many errors carry over from Faroese sound rules, and most do not reduce intelligibility. The authors argue that textbooks written for Danish or other Scandinavian learners do not suit Faroese students. They present the paper as a first step and call for further work on vowels and prosody.

Key points

• The six informants were one man and five women aged 25–39. They were recorded with Audacity and a Plantronics Audio 50 headset; the total was 3,547 analysed sound occurrences.
• Sampling per sound was capped as follows, giving at most 220 tokens per sound, and no sound was judged on fewer than 23 tokens:
  • up to 100 tokens from the key informant's reading;
  • up to 20 tokens from each of the other five readers;
  • up to 20 tokens from the key informant's free speech.
• /ʒ/ was the rarest sound. Approximants were analysed only from the key informant's reading.
• Problems with stops:
  • English /p t k/ come out with too little aspiration at the start and in the middle of words, which blurs pairs like pear–bear.
  • The same stops are pre-aspirated in the middle and at the end of words. The suggested fix at word ends is pre-glottalisation.
  • /b d g/ are devoiced too strongly, so learners should lengthen the vowel before a lenis stop.
  • /t d/ are dental, but this barely matters.
  • Intervocalic /t/ is sometimes voiced, which sounds more American than RP.
• Faroese /r/ is pronounced in every position, so learners should drop it in codas. Their /l/ is too dark and dental, and /l/ and /r/ may be devoiced before voiceless consonants.

Interesting ideas

• I1. Capping how many tokens each speaker contributes per sound, with a minimum token floor, stops one prolific source from dominating and flags sparse categories.
• I2. Checking one informant's spontaneous speech against their read speech, before trusting read data, is a cheap test for distribution shift.
• I3. Each target sound gets one "correct" category plus its own set of error types, rather than a shared generic error list.
• I4. Errors are split into phonemic, allophonic and distributional kinds, and each is also judged by whether it damages intelligibility or only sounds foreign.
• I5. Many learner errors are predicted by spelling out the native-language rule that transfers, such as sonorant devoicing before voiceless consonants.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Build evaluation sets with per-source caps and a minimum count per construction, and report categories below the floor as unreliable rather than scoring them.
• I2 (learn): Before trusting silver data from edited text, compare parser agreement on a small sample of informal text from the same domain, as a sanity control.
• I3 (tree): Give each UD relation its own error buckets, such as attachment-site mistakes for obl versus label confusions for nsubj/obj, so that analysis points to specific repair rules.
• I4 (back): Grade generated text by separating errors that change meaning from those that only hurt naturalness, and fix meaning-changing ones first in the text → graph → text compressor.
• I5 (tree): When converting a treebank from another UD dialect, list the source conventions explicitly so that systematic mismatches are predicted and rewritten by rule rather than discovered one by one.