README.md

metadata

language:
  - zh
  - en
license: apache-2.0
size_categories:
  - n<1K
task_categories:
  - automatic-speech-recognition
pretty_name: AASRBench
tags:
  - speech
  - audio
  - asr-evaluation

AASRBench

AASRBench is a bilingual audio benchmark for evaluating ASR output on spoken-language cleanup, correction, formatting, filtering, and rephrasing.

This repository contains the benchmark presented in AgenticASR: Refining Speech Recognition in Real-World Scenarios via an Agentic Approach.

Code: https://github.com/AnXMuy/AgenticASR

Contents

• 917 benchmark samples.
• 917 WAV files referenced by the benchmark records.
• 6,637 rubric questions across content, format, filter, and rephrase.
• 11 scenes: academic, customer_service, daily_chat, dictation_memo, explanation, meeting, navigation, passthrough, tech, vibe_coding, and voice_search.
• 510 Chinese samples and 407 English samples.

The 41 additional WAV files in the source working directory were marked obsolete and are not included in this release.

Files

data/
├── benchmark.jsonl
├── rubric.json
├── rubric.jsonl
└── audio/
    └── *.wav
metadata/
├── manifest.json
├── sha256sums.txt
└── validation-report.json

benchmark.jsonl contains one JSON object per line. Each record includes source_record_id, audio, oral, clean, scene, language, and voice metadata. The audio value is a path relative to data/, for example audio/academic__recheckjsonl10.wav.

rubric.json stores the rubric as an items array. rubric.jsonl stores the same rubric items as JSONL records. Benchmark and rubric records are joined by source_record_id.

Rubric Options

The rubric parser accepts these three equivalent encodings:

1. An array of objects with text and score fields.
2. An array of strings such as "pass (1)".
3. One comma-separated string such as "pass (1), fail (0)".

Both English and Chinese parentheses are accepted. The score categories are documented in metadata/manifest.json and the source benchmark statistics.

Validation

The release files are checked with the included SHA-256 manifest:

sha256sum -c metadata/sha256sums.txt

metadata/validation-report.json records the local structural validation result. The expected release scope is 917 benchmark records, 917 rubric records, and 917 referenced audio files.

License and Provenance

The code is released under the Apache License 2.0. Redistribution rights, audio provenance, annotation provenance, and privacy review should be confirmed by the dataset owner before downstream redistribution.

Repository Names

• Display name: AASRBench
• Hugging Face: Andrew0425/AASR-Bench
• ModelScope: MuyuanJ/AASR-Bench