biasinear 0.2.0

BiasInEar: Assessing Sensitivity in Audio Language Models Across Linguistic, Demographic, and Positional Variations

BiasInEar

Assessing Sensitivity in Audio Language Models Across Linguistic, Demographic, and Positional Variations

BiasInEar is a benchmark for evaluating speech bias in multilingual multimodal large language models (MLLMs). It provides 11,200 spoken multiple-choice questions across 3 languages, 7 accents, 2 genders, and 2 option orders.

Installation

# Core (metrics only)
pip install biasinear

# With data loading
pip install biasinear[data]

# With audio utilities
pip install biasinear[audio]

# With a specific model provider
pip install biasinear[gemini]     # Google Gemini
pip install biasinear[openai]     # OpenAI
pip install biasinear[nvidia]     # NVIDIA Build
pip install biasinear[mistral]    # Mistral

# Everything
pip install biasinear[all]

Or with uv:

uv pip install biasinear[all]

Model Providers

BiasInEar includes built-in support for several audio language model APIs:

Provider	Class	Install	Default Model
Google Gemini	GeminiModel	pip install biasinear[gemini]	gemini-2.5-flash
OpenAI	OpenAIModel	pip install biasinear[openai]	gpt-4o-audio-preview
NVIDIA Build	NvidiaModel	pip install biasinear[nvidia]	google/gemma-3n-e4b-it
Mistral	MistralModel	pip install biasinear[mistral]	voxtral-small-2507

API Keys

Set your API key as an environment variable:

export GEMINI_API_KEY="..."
export OPENAI_API_KEY="..."
export NVIDIA_API_KEY="..."
export MISTRAL_API_KEY="..."

Or pass directly when creating the model:

from biasinear.models import GeminiModel
model = GeminiModel(api_key="your-api-key")

Quick Example (Gemini)

from biasinear.models import GeminiModel
from biasinear.utils import concat_audio

model = GeminiModel()  # uses GEMINI_API_KEY env var
combined = concat_audio(question=q_bytes, options=opt_bytes)
output = model.generate(combined)
print(output["answer"], output["raw_response"])

See examples/ for complete provider scripts.

Quick Start

1. Load Data

from biasinear import load_dataset

# Load a specific config
dataset = load_dataset(config="en_Female")

# Load all configs merged
dataset = load_dataset()

2. Run Inference

from biasinear.utils import concat_audio
from biasinear.models import BaseModel

# Implement your model by extending BaseModel
class MyModel(BaseModel):
    def generate(self, audio: bytes) -> dict:
        # Your API call here
        return {"answer": "A", "raw_response": "..."}

model = MyModel("my-model")
output = model.generate(audio_bytes)

3. Evaluate

from biasinear import Evaluator

evaluator = Evaluator(
    predictions=["A", "B", "A", ...],
    references=["A", "A", "A", ...],
    question_ids=["q1", "q1", "q2", ...],
    groups={
        "language": ["en", "en", "zh", ...],
        "gender": ["Female", "Male", "Female", ...],
        "order": ["original", "reversed", "original", ...],
    },
)
results = evaluator.run()
# {
#     "accuracy": 0.75,
#     "entropy": {"mean": 0.32, "per_question": {...}},
#     "apes": {"language": 0.12, "gender": 0.03, "order": 0.15},
#     "fleiss_kappa": {"language": 0.65, "gender": 0.88, "order": 0.52},
# }

Use Metrics Individually

from biasinear import accuracy, question_entropy, apes, fleiss_kappa

acc = accuracy(predictions, references)
ent = question_entropy(["A", "A", "B", "C"], num_categories=4)
apes_val = apes([0.3, 0.5, 0.4])
kappa = fleiss_kappa(ratings_matrix)

Metrics

Metric	Description
Accuracy	Standard MCQ correctness
Question Entropy	Prediction uncertainty across configurations
APES	Average Pairwise Entropy Shift across variable levels
Fleiss' Kappa	Inter-rater agreement across perturbations

See the paper for details.

Citation

@inproceedings{wei-etal-2026-biasinear,
  title={Bias in the Ear of the Listener: Assessing Sensitivity in Audio Language Models Across Linguistic, Demographic, and Positional Variations},
  author={Wei, Sheng-Lun and Liao, Yu-Ling and Chang, Yen-Hua and Huang, Hen-Hsen and Chen, Hsin-Hsi},
  booktitle={Findings of the Association for Computational Linguistics: EACL 2026},
  year={2026},
  publisher={Association for Computational Linguistics}
}

License

Apache License 2.0. See LICENSE for details.