A post-ASR correction model trained on 90 million noisy / clean pairs, designed to fix typical speech recognition errors in (near) real-time. It helps clean up transcriptions from ASR systems like Whisper, Nemo etc, improving readability and grammatical correctness with minimal latency.

Designed to: Improve readability and professionalism in transcripts Make ASR outputs usable for customer service, legal, and healthcare Helping non-native speakers interpret ASR more easily Supports real-time captioning and assistive technologies

• Fixes common ASR issues:

  • Homophones (e.g., “their” vs “there”)
  • Subject-verb agreement
  • Verb tense errors
  • Missing auxiliaries and articles
  • Contractions and prepositions
  • Pronoun misuse
  • Repeated or corrupted words

• Fast inference (<50ms on GPU)

• Easy to plug into real-time ASR pipelines

• HuggingFace-compatible model loading

license: mit language:

• en base_model:
• google-t5/t5-small pipeline_tag: text2text-generation tags:
• ASR,

A lightweight grammar correction model fine-tuned from t5-small and t5-base, specifically designed to correct common errors in automatic speech recognition (ASR) outputs — including homophones, verb tense issues, contractions, duplicated words, and more. Optimized for fast inference in (near) real-time ASR pipelines.

• Small model: t5-small
• Base model: t5-base
• Fine-tuned on: 90 million synthetic (noisy → clean) sentence pairs
• Training objective: Correct ASR-style transcription errors into clean, grammatical English
• Framework: Hugging Face Transformers + PyTorch

1. Accuracy is a measure of how well the model performs across the full sentence. That is, a prediction is only counted as "correct" if the entire corrected sentence exactly matches the reference sentence. So if the model corrects 1 out of 2 errors, but the final output does not exactly match the expected sentence, it's counted as a fail.
2. Token Accuracy is a measure of how well the model performs at the token level. $\text{Token Accuracy (}=\left(\frac{\text{Number of Matched Tokens}}{\text{Total Reference Tokens}}\right)\times 100$

• Homophone mistakes (their → they're)
• Subject-verb disagreement (he go → he goes)
• Verb tense corruption (i seen → i saw)
• Missing auxiliaries (you going → are you going)
• Contraction normalization (she is not → she isn't)
• Repeated words (i i want → i want)
• Misused articles/prepositions/pronouns

Input (noisy ASR):

Models have been trained on DJ-AI Custom Dataset which includes over 90 million real and synthetic ASR errors and corrected texts pairs. The models are based on T5 pretrained models.

https://huggingface.co/dayyanj/dj-ai-asr-grammar-corrector-small

https://huggingface.co/dayyanj/dj-ai-asr-grammar-corrector-small-streaming

https://huggingface.co/dayyanj/dj-ai-asr-grammar-corrector-base

DEMO: https://huggingface.co/spaces/dayyanj/dj-ai-asr-grammar-corrector-demo

MIT License.