🏔️ ChuckleNet-Ten (天)

"The peak of laughter anticipation."

Human-verified laughter detection — trained on 87 Gillick videos with auditor-checked labels. Primary model for production use.

⚠️ Read this before evaluating: The headline "Train F1 0.975" is on 87 training videos. The number that matters for generalization is [email protected] = 0.330 on 118 held-out videos (comedian-disjoint split — Dave Chappelle held out, all others in training). See §Performance for the full breakdown.

---

What It Does

ChuckleNet-Ten detects laughter events in stand-up comedy audio using frozen WavLM embeddings fused with prosody features:

┌─────────────────────────────────────────────────────────────────┐
│                      ARCHITECTURE                                │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│   🎤 Audio Input (5s, 16kHz)     ┌─────────────────────┐      │
│                               ──►│  WavLM Backbone     │      │
│                                   │  microsoft/wavlm-base│     │
│                                   │  frozen, 768-dim    │      │
│                                   └──────────┬──────────┘      │
│                                              │                 │
│                                              ▼                 │
│                                   ┌─────────────────────┐      │
│   📊 Prosody (10→21 dim)    ────►│  Prosody Head      │      │
│   energy, F0, ZCR               │  BatchNorm + Drop  │      │
│                                   └──────────┬──────────┘      │
│                                              │                 │
│                                              ▼                 │
│                                   ┌─────────────────────┐      │
│                                   │  Fusion MLP        │      │
│                                   │  789→256→128→64→1 │      │
│                                   └──────────┬──────────┘      │
│                                              │                 │
│                                              ▼                 │
│                                   ┌─────────────────────┐      │
│                                   │  🤖 laugh / no_laugh│      │
│                                   └─────────────────────┘      │
└─────────────────────────────────────────────────────────────────┘

Key finding: Frozen WavLM already contains laughter-predictive information. A simple MLP probe + 10-dim prosody is sufficient to extract it.

---

Performance

Metrics on Held-out Test Set (Primary)

These are the numbers that matter. Evaluated on 118 Gillick videos from comedians not in training (Dave Chappelle held out).

Metric	Value	Note
[email protected]	0.3302	Primary metric — event-level overlap
Average Precision	0.2290	Precision-recall area on held-out
Training F1	0.975	On 87 training videos only — not a generalization number

Why the 66% Gap Between Train F1 and Held-out IoU-F1

Train F1 (87 videos):   ████████████████████████████ 0.975
Held-out IoU-F1 (118v): ████████░░░░░░░░░░░░░░░░░░░ 0.330
                                ↓ ~66% drop

The gap is real and expected:
• Training set: 87 videos, all same comedians
• Held-out set: 118 videos, DIFFERENT comedian (Dave Chappelle)
• Small-n training (87v) → overfitting to training distribution
• Distribution shift to unseen comedian is the main driver
• This is NOT a model failure — it's a fundamental small-dataset limitation

📌 Bottom line: Use [email protected] = 0.330 on held-out as the true performance estimate. Train F1 = 0.975 reflects label quality, not generalization.

Comparison: Weak Labels vs Human-Verified

              Weak Labels    Human-Verified    Δ
[email protected]:    0.229           0.330       +0.101 (+44%)
Training F1:   0.27            0.975       +0.705

Human verification improved held-out event detection by 44%.

---

Scientific Gates

The model passed 7 pre-registered scientific gates. Here's what each actually tests:

Gate	What It Tests	Result
G1	Labels independently grounded — are labels human-verified, not auto-generated?	✅ PASS
G2	Acoustic representation detects target — does WavLM carry any laughter signal?	✅ PASS (F1=0.33)
G3	Survives adversarial negatives — does model fail on new negative comedians?	✅ PASS (+0.11 on comedian-disjoint)
G4	Temporal context adds — does shuffled window order hurt?	✅ PASS (+0.077 true-order)
G5	Beyond transcript — do prosody features add F1 beyond text markers?	✅ PASS (+0.11 prosody-only)
G6	Not generic emotion — does model fail on MELD emotional speech?	✅ PASS (Δ=0.0 vs baseline)
G7	Genuine anticipation — are signals at ≥2.5s before onset real?	✅ PASS (+0.222 at ≥2.5s)

Full protocol: Das-rebel/autonomous_laughter_prediction

---

Beyond Transcript

The core scientific contribution: prosody features carry comedy timing information beyond what transcript markers provide.

Model Variant	Held-out IoU-F1	Δ from Baseline
WavLM + Prosody (no text)	0.330	baseline
Transcript-only baseline	~0.22	-0.11

Conclusion: Laughter anticipation works from prosody alone — words are not required.

---

Quick Start

# Install
pip install transformers torch torchaudio

# Use via pipeline (recommended — handles preprocessing)
from transformers import pipeline

pipe = pipeline(
    "audio-classification",
    model="Hayasuki/ChuckleNet-Ten",
    trust_remote_code=True   # required — model uses custom code
)

# Detect laughter
result = pipe("standup_clip.mp3")
# [{'label': 'laugh', 'score': 0.942}, {'label': 'no_laugh', 'score': 0.058}]

⚠️ trust_remote_code=True is required. This model ships custom Python code (ChuckleNetVerified) that is not part of the core transformers library.

---

Batch Processing Example

import os
from transformers import pipeline

pipe = pipeline(
    "audio-classification",
    model="Hayasuki/ChuckleNet-Ten",
    trust_remote_code=True
)

# Process all audio files in a directory
audio_dir = "path/to/comedy/clips"
results = []

for filename in os.listdir(audio_dir):
    if filename.endswith(('.mp3', '.wav', '.flac')):
        result = pipe(os.path.join(audio_dir, filename))
        results.append({
            'file': filename,
            'label': result[0]['label'],
            'confidence': result[0]['score']
        })

laughs = [r for r in results if r['label'] == 'laugh']
print(f"Laughter detected in {len(laughs)}/{len(results)} clips")

---

Training Details

Component	Specification
Backbone	microsoft/wavlm-base (frozen, 768-dim)
Prosody Features	10-dim: RMS energy ×5, F0 ×5
Classifier	MLP: 789→256→128→64→1
Optimizer	AdamW, lr=1e-3, weight_decay=0.01
Scheduler	CosineAnnealingLR, 50 epochs
Training Data	87 Gillick videos, human-audited labels
Hardware	NVIDIA GPU (CUDA)

---

Limitations

Limitation	Impact	Mitigation
Generalization gap (0.975→0.33)	Held-out performance is much lower	Use held-out IoU-F1=0.330 as true estimate
Single domain	English stand-up comedy only	Not tested on other genres
No fine-tuning	Frozen backbone limits adaptation	Consider fine-tuning last layers
5s windows	Long audio needs chunking	Use sliding window with stride

---

Related Models

Model	Labels	Held-out IoU-F1	Use Case
🏔️ ChuckleNet-Ten	Human-verified	0.330	Production ✅
👻 ChuckleNet-Kage	Weak (VTT)	0.229	Research / benchmarking
🎭 ChuckleNet-Genki	Weak (VTT)	~0.23	Historical reference

---

Resources

Resource	Link
Paper / Documentation	ChuckleNet-Ten Model Card (this page)
Research Repo	Das-rebel/autonomous_laughter_prediction
Dataset	StandUp4AI
Demo Space	laugh-prophecy
arXiv Preprint	Pending — cite this model card for now

---

Citation

For now, cite this model card. An arXiv preprint is in preparation.

@misc{chuckleNetTen2026,
  author = {Subhajit Das},
  title = {ChuckleNet-Ten: Verified Laughter Anticipation at Scale},
  year = {2026},
  url = {https://huggingface.co/Hayasuki/ChuckleNet-Ten},
  note = {Human-verified laughter detection using frozen WavLM + prosody fusion. 
          Held-out [email protected] = 0.330 on 118 comedian-disjoint videos.}
}

---

Model Metadata (LLM-Parseable)

{
  "model_name": "ChuckleNet-Ten",
  "model_name_zh": "笑音Net-巅",
  "architecture": "frozen_wavlm_base_plus_mlp_with_prosody",
  "primary_task": "audio_classification",
  "secondary_task": "laughter_detection",
  "input_modality": "audio",
  "input_format": "audio_waveform_16kHz_mono",
  "output_format": "binary_classification",
  "labels": ["laugh", "no_laugh"],
  "training_data": "87_Gillick_videos_human_verified",
  "held_out_data": "118_Gillick_videos_dave_chappelle_disjoint",
  "held_out_iou_f1": 0.3302,
  "held_out_ap": 0.229,
  "training_f1": 0.975,
  "key_findings": [
    "frozen_wavlm_contains_laughter_predictive_signals",
    "prosody_features_carry_comedy_timing_beyond_transcript",
    "laughter_anticipation_5_to_15_seconds_before_onset",
    "human_verification_improves_iou_f1_by_44_percent"
  ],
  "use_cases": [
    "real_time_comedy_performance_analysis",
    "audience_response_prediction",
    "comedy_script_optimization"
  ],
  "limitations": [
    "generalization_gap_0.975_train_to_0.330_held_out",
    "single_domain_english_standup_comedy",
    "no_backbone_fine_tuning"
  ]
}

---

Part of the ChuckleNet laughter detection family — detecting comedy before it happens.