Loading
Tem um workflow do ComfyUI? Gere uma documentação legível direto no navegador.Torne um workflow legível.Experimentar agora ↗
Ostrich-v1 is a production-grade multi-label audio classification model fine-tuned for detecting musical moods in Indian music (Bollywood, Indie, Folk, Classical, Fusion, etc.). It uses a teacher–student distillation...
Fonte do modelo
Descrição da fonte
Ostrich-v1 is a production-grade multi-label audio classification model fine-tuned for detecting musical moods in Indian music (Bollywood, Indie, Folk, Classical, Fusion, etc.). It uses a to map rich acoustic representations to , including culturally specific emotional categories.
Fontes
1 fonteVerificado 19 de set.
Artefatos de modelo
1 artefatoTrechos de fonte
2 trechos
sandychoii/distilhubert-finetuned-gtzan-audio-classificationOstrich-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:
The output is a 31-dimensional probability vector, allowing multiple moods to coexist per track.
beastLucifer/ostrich-v1-audio-moodbeastLucifer/sangeetkar-mood-datasetbeastLucifer/sangeetkar-teacher-labelsLabel 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:
Temporal Assumption:
Optimized for ≤30s chunks; long-form compositions should be chunked.
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))
}
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:
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
beastLucifer
--- library_name: transformers tags: - audio-classification - mood-detection - music - pytorch - hubert license: apache-2.0 datasets: - beastLucifer/sangeetkar-mood-dataset - beastLucifer/sangeetkar-teacher-labels language: - en metrics: - accuracy base_model: - sandychoii/distilhubert-finetuned-gtzan-audio-classification pipeline_tag: audio-classification --- # Ostrich-v1: Audio Mood Classification Model ## Model Overview **Ostrich-v1** is a production-grade **multi-label audio classification model** fine-tuned for detecting **musical moods in Indian music** (Bollywood, Indie, Folk, Classical, Fusion, etc.). It uses a **teacher–student distillation pipeline** to map rich acoustic representations to **31 fine-grained mood dimensions**, including culturally specific emotional categories. - **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...
Source context: 8 downloads · 0 likes · Pipeline audio-classification · Library transformers · Repo beastLucifer/ostrich-v1-audio-mood