Ostrich v1 Audio Mood

• Developed by: beastLucifer
• Model type: Multi-label Audio Classification
• Domain: Music (Indian & South Asian context)
• Languages: English, Hindi (musical semantics)
• License: Apache 2.0
• Finetuned from: sandychoii/distilhubert-finetuned-gtzan-audio-classification

Model Description

Ostrich-v1 bridges the gap between generic audio emotion models and the emotional textures unique to South Asian music.
The model is built on a DistilHuBERT backbone, fine-tuned on the Sangeetkar dataset using teacher-generated soft labels produced by a CLAP-based audio–text model.

It supports both:

• Global moods (e.g., Happy, Sad, Angry)
• Regionally grounded moods (e.g., Dard-bhari, Tapori, Sufi-romantic)

The output is a 31-dimensional probability vector, allowing multiple moods to coexist per track.

Model Sources

• Repository: beastLucifer/ostrich-v1-audio-mood
• Training Audio Dataset: beastLucifer/sangeetkar-mood-dataset
• Teacher Label Dataset: beastLucifer/sangeetkar-teacher-labels

Intended Uses

Direct Use

• Automated music tagging and metadata enrichment
• Mood-based playlist generation
• Music discovery systems for Indian sub-genres
• Recommendation systems and catalog analytics

Out-of-Scope Use

• Speech-to-text or speaker recognition
• Environmental sound classification
• Real-time or ultra–low-latency streaming inference
• Non-musical audio domains

Bias, Risks, and Limitations

• Label Noise:
  Labels are distilled from a teacher model. Although class-wise weighting is applied, subtle secondary moods may bleed into primary predictions.

• Genre Bias:
  Performance may degrade on:

  • Purely instrumental tracks
  • Rare regional folk styles underrepresented in the teacher corpus

• Temporal Assumption:
  Optimized for ≤30s chunks; long-form compositions should be chunked.

How to Use the Model

from transformers import AutoModelForAudioClassification, AutoFeatureExtractor
import torch
import librosa

model_id = "beastLucifer/ostrich-v1-audio-mood"

model = AutoModelForAudioClassification.from_pretrained(model_id)
processor = AutoFeatureExtractor.from_pretrained(model_id)

# Load audio
audio, sr = librosa.load("your_song.wav", sr=16000)

inputs = processor(
    audio,
    sampling_rate=16000,
    return_tensors="pt"
)

# Inference
with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.sigmoid(logits)  # multi-label probabilities

# Map predictions
id2label = model.config.id2label
predictions = {
    id2label[i]: probs[0][i].item()
    for i in range(len(id2label))
}

Training Details

Training Data

• Audio: Large-scale Indian music corpus
• Labels: 31 mood dimensions generated via a CLAP-based teacher model

Preprocessing

• Resampling: 16 kHz
• Normalization: Zero-mean, unit-variance
• Chunking: Max 30 seconds per sample

Training Configuration

• Optimizer: adamw_bnb_8bit

• Learning Rate: 5e-5

• Batch Size: 4 per device

• Effective Batch Size: 16 (gradient accumulation)

• Precision: FP16 mixed precision

• Loss Function: Custom Weighted BCEWithLogitsLoss

• Special Handling:

  • Motivational (index 19) weight reduced to 0.3 due to high teacher variance

Architecture

• Backbone: DistilHuBERT
• Objective: Multi-label mood classification
• Distillation: Teacher–student training for compactness and speed
• Inference: ~90% of HuBERT performance at significantly reduced compute cost

Compute Infrastructure

Hardware

• NVIDIA Tesla T4 (Google Colab)

Software

• PyTorch
• Hugging Face Transformers
• Accelerate
• BitsAndBytes (8-bit optimization)

Label Inventory (31 Classes)

Energetic Calm Happy Sad Angry Romantic Mysterious Nostalgic Dard-bhari Masti Sufi-romantic Item Song Qawwali Vibes Judaai Tapori Chill-lofi Hype Party Dreamy Dark Motivational Melancholic Intense Peaceful Experimental Ambient Spiritual Groovy Folk Indie Electronic Classical

Model Card Authors

beastLucifer