cadenzachallenge/CLIP2-BaselineMono

Baseline for the Second Cadenza Lyric Intelligibility Prediction (CLIP2) challenge

The baseline model for the CLIP2 Cadenza challenge consisting of a regression model that predicts a scalar in [0, 1] representing how intelligible a song's lyrics are. Built on a partially fine-tuned Whisper encoder as front-end and a small MLP regression head.

The lyric_intelligibility_model.py is self-contained: it has the model (WhisperIntelligibilityModel, with save_pretrained()/from_pretrained() mirroring the standard HF API) and a ready-to-use inference wrapper (IntelligibilityPredictor) in one file.

Files in this repo

File	Purpose
config.json	Model hyperparameters (backbone, hidden_dim, layer selection)
model.safetensors	Trained weights (state_dict)
lyric_intelligibility_model.py	WhisperIntelligibilityModel + IntelligibilityPredictor — everything needed to load and run the model
inference.py	Optional CLI wrapper around IntelligibilityPredictor that writes results to a CSV file

Loading and using the model

from lyric_intelligibility_model import IntelligibilityPredictor

predictor = IntelligibilityPredictor.from_pretrained(
    "cadenzachallenge/CLIP2-BaselineMono"
)
result = predictor.predict("path/to/song.wav")
print(result["score"])          # e.g. 0.83
print(result["channel"])        # "mono", or "left"/"right" for stereo (better-ear)

predict() also accepts a directory: it scores every audio file at the top level (not recursive) and returns {filename: result} instead of a single result dict.

results = predictor.predict("path/to/songs_dir")
for filename, r in results.items():
    print(filename, r["score"])

Stereo audio is automatically scored per-channel and combined via the better-ear strategy (max(score_left, score_right)) by default — pass better_ear=False to downmix to mono instead. Audio is resampled to 16 kHz internally using soxr (SoX HQ).

If you just want the raw model (e.g. for further fine-tuning) without the audio-loading wrapper:

from lyric_intelligibility_model import WhisperIntelligibilityModel
from transformers import WhisperProcessor

model = WhisperIntelligibilityModel.from_pretrained(
    "cadenzachallenge/CLIP2-BaselineMono"
)
processor = WhisperProcessor.from_pretrained(
    model.SUPPORTED_MODELS['whisper-large-v3']
)

# unfreeze some encoder layers for fine-tuning
model.unfreeze_backbone(num_top_layers=4, context_layers=2)

# preprocess audio -> log-mel features
inputs = processor(waveform, sampling_rate=16000, return_tensors="pt")
scores = model(inputs.input_features)  # forward pass, then compute your loss

Command-line usage

inference.py is a thin CLI wrapper around IntelligibilityPredictor that writes a results.csv (columns: filename stem, score — no header) instead of returning a dict:

pip install torch transformers huggingface_hub safetensors soundfile soxr

python inference.py \
    --repo_id cadenzachallenge/CLIP2-BaselineMono \
    --audio path/to/song.wav          # or a directory

# options:
#   --no-better-ear   downmix stereo to mono instead of left/right max
#   --max-100         scale scores to [0, 100] instead of [0, 1]
#   --output FILE     custom CSV path (default: results.csv)

Model details

• Backbone: Whisper encoder (size set by config.json, default whisper-large-v3)

• Layer fusion: mean-pooled encoder layer(s) → learned per-layer weights → optional cross-layer attention (when multiple layers are selected)

• Head: 2-layer MLP + sigmoid → score in (0, 1)

• Output: single scalar intelligibility score per input

• Data: The model was trained in CLIP2 train dataset.

More Information

For more information, please see the Cadenza Challenge website.