Audiction
a hybrid text normalizer for text-to-speech systems
2 min readupdated
Audiction converts written English text into a spoken-word format ready for Text-to-Speech (TTS) engines—turning abbreviations, dates, currency amounts, and numbers into fully spelled-out words (for example, "11/21" becomes "November twenty-first"). Developed by OmniSpace Technologies, it builds on the google/byt5-small model (~300M parameters) paired with deterministic linguistic rules. It is built for voice assistant developers, podcasters, and content creators using TTS pipelines who want to avoid mispronounced numbers, acronyms, and special characters.
Key features
- Hybrid pipeline: hand-crafted rules handle standard patterns in microseconds, while the AI model only kicks in for complex or ambiguous cases.
- High accuracy: reaches 99.9% and higher success rates across both synthetic and real-world test sets.
- Clean training data: trained on synthetic templates and a subset of the FineWeb dataset, with no proprietary or commercial texts used.
- Permissive Apache 2.0 license allows commercial use and modification.
What you can do
- Place the normalizer ahead of any TTS model in a bot or voiceover service to read dates, prices, and addresses properly.
- Prep radio scripts and announcements by automatically expanding digits into spoken words without manual edits.
- Call the hybrid pipeline from Python to clean up complex documents or logs before speech synthesis.
Example
from normalize import Normalizer
norm = Normalizer(".") # hybrid: rules first, the model only where the rules refuse
norm(["Dr. Patel joins us from Pine Ave. after the break."])
Read more
DetailsBenchmarks: 1.000 on 2,000 synthetic samples (eval_synth.jsonl), 0.9995 on eval_synth_v3.
Hardware: trained on Modal H100s; inference runs on standard CPUs or GPUs.
Limitations: tailored to US English conventions (dates, US dollars, Fahrenheit) and does not handle URLs, email addresses, or promo codes.










Comments