Audiction

Audiction: a 1920s robot announcer at a broadcast microphone

Audiction turns written English into the words a text-to-speech voice should say. Put it in front of any TTS voice: radio scripts, assistants, LLM output.

How to run

pip install "transformers>=4.46,<5" torch huggingface_hub
hf download Frosty40/Audiction --local-dir audiction && cd audiction
python normalize.py "Doors open at 7:30 pm on Nov 21 at 135 Prospect St."

From Python:

from normalize import Normalizer
norm = Normalizer(".")
norm(["Call 1-800-555-0199 or email news@example.org."])

Examples

Written Spoken
Doors open at 7:30 pm on Nov 21 at 135 Prospect St. Doors open at seven thirty p m on November twenty-first at one hundred thirty-five Prospect Street.
Dr. Patel joins us from Pine Ave. after the break. Doctor Patel joins us from Pine Avenue after the break.
Louis XIV built the palace in the 1660s. Louis the Fourteenth built the palace in the sixteen sixties.
Add 2 1/2 cups of flour and bake at 350 °F. Add two and a half cups of flour and bake at three hundred fifty degrees Fahrenheit.
The NBA draft starts at 8 pm ET. The N B A draft starts at eight p m E T.
Call 1-800-555-0199 or email news@example.org. Call one, eight hundred, five five five, zero one nine nine or email news at example dot org.
Stream it in 1080p at radio.example/live. Stream it in ten eighty p at radio dot example slash live.
The Sooners won 35-28 on 2026-11-21. The Sooners won thirty-five to twenty-eight on November twenty-first twenty twenty-six.
Use code SAVE20 for a $5 discount. Use code SAVE twenty for a five dollar discount.
The fund raised £100m, up 12.5% on last year. The fund raised one hundred million pounds, up twelve point five percent on last year.
A dimmable MR16 lamp draws 7W. A dimmable M R sixteen lamp draws seven watts.

What it reads

  • Money: dollars, pounds and euros, cents, "$5bn", "£100m", and money used as an adjective ("a five dollar discount").
  • Time and dates: clock times, "10.30am", "Nov 21", "11/21", "2026-11-21", "28/06/2017", years and decades.
  • Numbers: ordinals, ranges and scores, percentages, decimals, fractions and mixed numbers.
  • Units: units of measure, degrees, coordinates.
  • Contact details: phone numbers (including 1-800 and 911), ZIP codes, web addresses and email addresses.
  • Words and codes:
    • abbreviations (Dr., St., Ave., Jr., No., vs.)
    • Roman numerals (Louis XIV, Super Bowl LX)
    • acronyms said letter by letter (NBA, FBI, ET)
    • product and model codes (MR16, F-150, PS5, 1080p)
    • race distances (10K)

It's written for US English: dollars, °F and month/day dates.

How it works

  1. The rules in verbalize.py speak a sentence when they're sure of every part. They're exact and take microseconds.
  2. The model (google/byt5-small, fine-tuned) reads the sentences the rules leave open. Its reading is used only when it passes a check:
    • the sentence's own words are kept, in order
    • every word it adds is a spoken-form word (number words, units, months)
    • every unit and currency word matches a symbol in the source
    • no digit is left
  3. Otherwise, the rules speak everything they're sure of, and the rest stays as written.

Results

Exact match of the full pipeline (eval_hybrid.py) on held-out sentences built from templates that training never saw. The spoken forms follow the conventions above.

Eval Pairs Exact match
eval/eval_synth.jsonl: money, times, dates, numbers, units 2,000 1.000 (2,000/2,000)
eval/eval_synth_v4.jsonl: fractions, dates, titles and streets, Roman numerals, acronyms, race distances 2,000 1.000 (2,000/2,000)

Hand-written radio lines (eval_real.py), written after this release and scored once. Scoring ignores capitals, since a voice can't hear them:

Eval Lines Correct
eval/eval_real_v2.jsonl 60 59/60

Run it yourself: python eval_hybrid.py --eval eval/eval_synth.jsonl or python eval_real.py --model . --eval eval/eval_real_v2.jsonl.

Training

  • Model: google/byt5-small, fine-tuned for 3 epochs: 12,195 steps at effective batch 64, on one H100.
  • Data: about 60k synthetic and 200k FineWeb sentence pairs per epoch. Every target is spoken by the rules and checked so that it only uses source words and spoken-form words.

Data

  • Synthetic pairs: sentences from templates we wrote, filled with random values. There's no outside text in them, so no outside licence.
  • Web pairs: sentences from FineWeb sample-10BT (ODC-By 1.0), spoken by the rules. Thanks to the FineWeb team.
  • No Wikipedia text and no non-commercial data.

Licence

Apache 2.0, the same as the base model google/byt5-small. You can use, change and redistribute it, commercially too, as long as you keep the licence and notices. The FineWeb sentences used in training are ODC-By 1.0 and credited above.

Banner art was made for this project with an AI image model.

Built by OmniSpace Technologies. More OmniSpace work: huggingface.co/Omnispace

Downloads last month
202
Safetensors
Model size
0.3B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Frosty40/Audiction

Finetuned
(348)
this model

Dataset used to train Frosty40/Audiction