aguangguang/LCAR-Hallucination-Benchmark

LCAR Hallucination Benchmark

LCAR Hallucination Benchmark is a manually reviewed speech benchmark for studying acoustic-grounding failures in LLM-based ASR. It contains two 500-utterance suites: controlled speech synthesized with IndexTTS2 and speech derived from openly released corpora. The benchmark covers translation or transliteration, spoken or text-prompt instruction execution, unsupported repetition, and catastrophic deletion.

The benchmark is a targeted stress set. It is intended to evaluate whether a decoder leaves a hallucinated trajectory, not to estimate hallucination prevalence in ordinary speech or compare the general recognition accuracy of the source ASR systems.

Repository Structure

LCAR-Hallucination-Benchmark/
|-- README.md
|-- LICENSE
|-- tts/
|   |-- indextts2.jsonl
|   `-- wav/
|       `-- *.wav
`-- openspeech/
    |-- openspeech.jsonl
    `-- wav/
        `-- *.wav

Dataset Summary

Suite	Utterances	Description
IndexTTS2	500	Controlled TTS speech with coherent Mandarin-English content and targeted prompting conditions
OpenSpeech	500	Speech derived from open-source Mandarin-English corpora with source provenance retained

Record Format

Both metadata files use JSON Lines. Each line describes one audio sample.

Field	Description
id	Stable release identifier
audio	Relative path to the WAV file
suite	IndexTTS2 or OpenSpeech
hallucination_type	Manually accepted failure category
reference	Reference transcript
prompt	Adversarial text prompt for Prompt-HAL; empty otherwise
hypothesis	Baseline ASR output exhibiting the failure
source	OpenSpeech source provenance; omitted from the TTS records

Loading

import json
from pathlib import Path

root = Path("LCAR-Hallucination-Benchmark")
with (root / "tts" / "indextts2.jsonl").open(encoding="utf-8") as f:
    records = [json.loads(line) for line in f]

audio_path = root / "tts" / records[0]["audio"]

License and Attribution

This repository contains material governed by multiple upstream licenses and therefore uses the Hugging Face other license tag. IndexTTS2-generated audio is subject to the bilibili Model Use License Agreement. OpenSpeech records retain source-level provenance and remain subject to the terms of their respective upstream datasets. See LICENSE before using or redistributing the benchmark.

Citation

Citation metadata will be added with the accompanying paper release.