Implements the LIR-ASR correction paradigm proposed in: 📄 “Listening, Imagining & Refining: A Heuristic Optimized ASR Correction Framework with LLMs” (arXiv:2509.15095)

• 📂 Datasets
• 📊 Evaluation Metrics
• ⚙️ Supported ASR Engines
• 🧩 LIR-ASR Correction Framework
• 🚀 Usage
• 📚 References

The framework supports the following datasets (and can be extended easily):

• LibriSpeech (test-clean / test-other)
• TED-LIUM
• CommonVoice
• Multilingual LibriSpeech (MLS)
• VoxPopuli
• Fleurs (multi-language, with helper script scripts/download_fleurs.sh)

We evaluate ASR and correction performance using:

• 📝 Word Error Rate (WER) — word-level edit distance / reference count
• ✒️ Punctuation Error Rate (PER) — errors in ., ,, ?, etc.
• ⏱️ Core-Hour — CPU hours per 1h audio (for local models)
• 💾 Model Size — size (MB) of acoustic + language models

⚠️ For cloud ASR services, Core-Hour and Model Size are not reported.

LIR-ASR (Listening → Imagining → Refining) is an iterative correction framework inspired by how humans “rehear” ambiguous speech.

Pipeline:

1. 👂 Listening — detect uncertain/misrecognized words
2. 💭 Imagining — generate candidate variants (phonetic substitutions, G2P)
3. ✨ Refining — use LLMs / scoring models to pick the most consistent

Key Features:

• 📌 FSM Controller — 3 states: NoSearch → Search → Search++
• 🧭 Heuristic Optimization — rule-based semantic constraints
• 🔄 Iterative Refinement — until convergence, ensuring monotonic score improvements

✅ Achieves ~1.5% CER/WER improvements on average over uncorrected baselines.

# Install dependencies
pip3 install -r requirements.txt

# Prepare datasets
sh scripts/download_fleurs.sh  # example for Fleurs

from optim import prompt_optimization, evolutionary_prompt_optimization, nbest_optimization, RIF_ASR
from normalizer import Normalizer
from languages import Languages

# Init normalizer
norm = Normalizer.create(language=Languages.ZH, keep_punctuation=True, punctuation_set=".?")

asr_text = "由于分离和重组便宜在每一代的两个库之间来回变动"

# 1. Prompt optimization
corrected = prompt_optimization(asr_text, llm="Qwen3-235B")

# 2. Evolutionary optimization
refined = evolutionary_prompt_optimization(asr_text, llm="Qwen3-235B")

# 3. RIF-ASR with multiple candidates
result = RIF_ASR(asr_text, llm="Qwen3-235B", language="ZH", normalizer=norm)

sh scripts/evaluation_whisper_qwen.sh

@misc{liu2025listeningimaginingrefining,
      title={Listening, Imagining & Refining: A Heuristic Optimized ASR Correction Framework with LLMs}, 
      author={Yutong Liu and Ziyue Zhang and Cheng Huang and Yongbin Yu and Xiangxiang Wang and Yuqing Cai and Nyima Tashi},
      year={2025},
      eprint={2509.15095},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2509.15095}, 
}