Paired (canonical G2P IPA, real ASR-output IPA) for the same 52,855 utterances,
for training a bidirectional noise-channel model:
noisify: true_ipa -> asr_ipa, predicting what
ghananlpcommunity/ghana-english-phoneme-asr
will actually output given clean text (including that it drops punctuation
entirely), for augmenting ipa2text training data with realistic - not
hand-written synthetic - noise.
denoise: asr_ipa -> true_ipa, cleaning real ASR output and… See the full description on the dataset page:
https://huggingface.co/datasets/ghananlpcommunity/true-asr-ipa-pairs.
