SpeechJudgeAudit

SpeechJudgeAudit tests whether speech LLM judges systematically prefer surface acoustic cues such as louder speech, richer or longer content, repetition, silence padding, and emotional delivery.

The release defines 2,425 speech assets. It packages 1,625 WAV files; the remaining 800 emotion files are reconstructed from ESD and are not redistributed.

Quick Start

Download the benchmark, then run preparation, task generation, your judge, and analysis with one command:

git clone https://huggingface.co/datasets/mingyue66/SpeechJudgeAudit SpeechJudgeAudit
cd SpeechJudgeAudit

bash run_benchmark.sh --from-hf --runner /path/to/judge_adapter

On first use, the script installs benchmark dependencies into an isolated .benchmark_env. It uses the minimal templates in prompts/ unless custom prompt files are supplied. The executable adapter receives two arguments: the generated task JSONL and the path where it must write normalized model outputs; BENCHMARK_ROOT points to the audio root. Without --runner, the script prepares the complete benchmark and stops after generating outputs/eval_items.jsonl.

Contents

Subset	Speech assets	Description
loudness/	400	Reference and +/-3 and +/-6 dB variants
content_richness/	298	Concise and richer meaning-equivalent speech
repetition_controls/	447	Concise, repeated, and duration-control speech
silence_padding/	480	Silence-padded clips and AB/BA concatenated comparisons
emotion/	800	Neutral, Happy, Sad, and Angry ESD speech

One additional separator-tone WAV is marked role=utility_separator and is never evaluated as speech. The packaged TTS audio is mono 24 kHz WAV; prepared ESD audio is mono 16 kHz PCM16. Use an audio loader that supports both WAV encodings.

How It Works

1. Prepare ESD

The pipeline selects the exact 20 ESD sentences used in the paper, verifies their transcripts, and reproduces the trimming and -23 LUFS normalization. Source positions and expected hashes are listed in metadata/esd_required_files.csv.

bash scripts/prepare_esd.sh --from-hf
# or
bash scripts/prepare_esd.sh --esd-root /path/to/ESD

2. Adapt Your Judge

Use the judge's official prompt, quality rubric, scoring scale, chat template, and output schema as closely as possible. Do not reveal the manipulated condition. Minimal templates are provided in prompts/.

Transcripts are optional auxiliary metadata. Pass them only when the judge normally accepts text; audio-only judges should ignore all transcript fields.

The adapter reads each generated task, attaches the WAV file through the judge's native processor or API, and writes one normalized result per line:

{"task_id": "...", "label": "A"}

Pairwise labels are A, B, or tie. A pointwise adapter instead writes {"task_id": "...", "score": 4.2}.

To replace the minimal templates with prompts tailored to a specialized judge:

bash run_benchmark.sh --from-hf \
  --runner /path/to/judge_adapter \
  --pairwise-prompt /path/to/pairwise_prompt.txt \
  --concatenated-prompt /path/to/concatenated_prompt.txt

3. Generate Tasks

The default command generates 3,054 counterbalanced pairwise trials:

python scripts/build_eval_jsonl.py --manifest manifest.csv --output eval_items.jsonl

Add --include-pointwise to also generate 2,105 pointwise trials. Supply --pointwise-prompt /path/to/pointwise_prompt.txt when the judge requires a custom pointwise rubric. The output is a model-independent task specification, not a universal API payload; the model adapter performs the final conversion.

4. Analyze Results

python scripts/analyze_results.py \
  --items eval_items.jsonl \
  --results model_outputs.jsonl \
  --output analysis_summary.csv

For pairwise trials, preference rates above 50% indicate preference for the manipulated condition. For pointwise trials, a positive mean score difference indicates that the manipulated condition received higher scores than its matched reference or control. The analyzer validates result coverage by default; use --allow-partial only for diagnostic runs. These quantities are audit diagnostics, not ground-truth accuracy scores.

Reference Baselines

Pairwise preference rates from the paper are shown below. Ties count as 0.5. Richer compares richer with concise speech; Longer (padding) compares silence-padded with concise speech, changing duration without adding content. Each emotion is compared with matched Neutral speech.

Judge	+6 dB	Richer	Longer (padding)	Happy	Sad	Angry
SpeechJudge-GRM	54.1%	57.0%	52.7%	47.8%	39.2%	39.5%
UniSRM	57.5%	84.6%	8.2%	59.8%	55.5%	53.5%
SQ-LLM	59.1%	75.5%	42.4%	55.1%	32.8%	47.3%
Gemini-2.5-Pro	56.6%	66.4%	50.7%	47.0%	36.6%	42.9%
Gemini-3.1-Pro	61.6%	70.1%	50.2%	55.8%	37.2%	48.2%

Across rationale-producing judges, direct cue naming was below 1% for loudness, 1.2% for silence padding, at most 3% for duration and emotion, and 22.1% for repetition.

Citation

@misc{acoustic_shortcuts,
  title={Louder, Longer, Livelier: Acoustic Shortcuts and Underspecified Rationales in Speech LLM Judges},
  author={Huo, Mingyue and Mehta, Shivam and Jawade, Bhavin and Lan, Yinghong and Li, Haoqi},
  year={2026},
  note={Preprint}
}

License and Sources

• This benchmark repository is released under CC BY-NC 4.0; upstream terms and required attributions continue to apply.
• Text and speech materials derive from LibriTTS and VCTK (CC BY 4.0), ASSET (CC BY-NC 4.0), and ESD (non-commercial research use). ESD audio is not redistributed.
• Generated stimuli use CosyVoice2 and Qwen3-TTS (Apache 2.0), plus F5-TTS and StyleTTS2 (MIT). No model weights are included.