hem-flan-t5-small

Turns clean scripted text into realistically disfluent speech, on a four position dial. It is the inverse of daltonoscar0/mend-flan-t5-small, which removes fillers and self-repairs from dictation.

Code, evaluation and the measurements behind the dial: github.com/daltonoscar0/hem

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("daltonoscar0/hem-flan-t5-small")
tok = AutoTokenizer.from_pretrained("daltonoscar0/hem-flan-t5-small")

text = "Send the quarterly report to Sarah before Friday."
ids = tok("<d2> say aloud: " + text, return_tensors="pt")
print(tok.decode(model.generate(**ids, max_new_tokens=192)[0], skip_special_tokens=True))

The input format is not optional: a control token, then say aloud: , then the clean written sentence. The four control tokens were added to the tokenizer as special tokens during fine-tuning, so a tokenizer loaded from anywhere else will split them into pieces and the dial will not work.

The dial

Token	Setting	Target rate
<d0>	clean	nothing inserted
<d1>	light	6.7 events per 100 words, one per 15
<d2>	natural	14.2 events per 100 words, one per 7
<d3>	heavy	28.1 events per 100 words, one per 4

The rates are not chosen. They are measured off the Switchboard Dialog Act corpus: <d1> is the median disfluent utterance and below, <d2> runs from the median to the 90th percentile, <d3> is the 90th percentile and above. The per-type breakdown and how it was counted are in the repository README.

Output register

Output is lowercased with punctuation stripped, which is the register Mend reads. That is deliberate: the two models are inverses and their output and input formats have to match.

Training

flan-t5-small fine-tuned on 50,000 synthetic pairs from a rule-based Shriberg-structured injector plus 24,619 real Switchboard utterances read backwards (33% real), 2 epochs, batch 32, learning rate 5e-5, greedy decoding. Every pair is labelled by its own measured disfluency rate rather than by the setting it was generated at.

The Switchboard split is by conversation, so the conversations the model was evaluated on were never trained on.

Limitations

Read the repository README's "Where it fails" section before using this. In short: the register is American telephone conversation, so the filled pauses it inserts can be wrong for a formal script; it does not fix agreement or determiners across a repair it has just created; restarts are the type it handles worst; and it will occasionally drop or alter a word of the script rather than only adding to it, which matters if the script is something someone has to say verbatim.

Licence

MIT.