A Core ML build of Parakeet Ultra whose decoder and joint were fine-tuned to write numbers, symbols and
abbreviations the way they are spoken ("twenty five percent", "point nine", "December thirty first") instead of
mixing digits and words. The encoder, preprocessor, tokenizer and file layout are unchanged from
FluidInference/parakeet-ultra-coreml, so it is a
drop-in replacement: only Decoder.mlmodelc/weights/weight.bin and the joint weights
(JointDecision.mlmodelc / JointDecisionv3.mlmodelc, byte-identical) differ.
Built for MimicScribe, which transcribes overlapping audio windows and stitches them;
one consistent written form for numbers makes the overlaps agree.
Lineage and attribution
Training data
- Synthetic speech generated with Kokoro (Apache 2.0) from templated
sentences containing numbers, dates, money, codes and abbreviations, with spoken-form labels
- LibriSpeech-PC (CC BY 4.0)
- FLEURS (CC BY 4.0)
- ICSI Meeting Corpus (CC BY 4.0) — word-labelled meeting crops
The real-speech data is replayed with a distillation term toward the original model, to keep its recognition and
punctuation behaviour on everything except written form.
Results
Full MimicScribe post-recording pipeline, same build, 27 public meetings (16 AMI + 11 Earnings-21, 18.8 h):
|
Parakeet Ultra |
This model |
| WER, content words, numbers excluded |
12.85% |
12.84% |
| WER, verbatim (fillers counted) |
16.95% |
16.47% |
| Numbers read exactly (Earnings-21) |
96.8% |
96.8% |
Recognition is on par; the change is the written form, which is now spoken words throughout. Years keep both
halves ("twenty twenty one"): 2 of 138 held-out Earnings-22 years collapse ("two twenty one"). Expect slightly more
sentence breaks on conversational speech: on held-out Earnings-22 clips, sentence-end F1 is 84.9 against the base
model's 89.8.
Use
Same inputs, outputs and decoding as Parakeet Ultra / v3 TDT in FluidAudio. Text comes out in spoken form; apply
your own inverse text normalization if you want digits.
