Apppublic

explorer13/risk-level-classification

sourceHugging Facemitupdated 4mo agoView on Hugging Face

0likes

App README

Risk Level Classification API

A real-time audio risk assessment API built for women's safety applications. When an SOS is triggered, the app captures audio and sends it here to determine the threat level.

How It Works

text

🎙️ Audio (WAV, ≤35s)
    │
    ▼
🎧 Whisper (language detection from audio signal)
    │
    ▼
🗣️ Google STT (transcription in native script — Bengali, Devanagari, etc.)
    │
    ▼
🔤 Romanization (native script → Latin characters)
    │
    ▼
🌐 Translation (→ English)
    │
    ▼
🤖 Risk Classification (Low / Medium / High + confidence score)

Key design decision: Whisper detects the language from audio acoustics (never confuses Hindi/Bengali), then Google STT transcribes using that language code (produces clean native script). Best of both worlds.

Quick Start

bash

curl -X POST "https://your-space.hf.space/analyze_audio" \
  -F "file=@recording.wav"

Example Response

json

{
  "original_text": "আমাকে আমার মত থাকতে দাও",
  "romanized_text": "amake amar moto thakte dao",
  "translated_text": "Let me be who I am",
  "risk_level": "Low",
  "score": 0.77
}

Running Locally

bash

pip install -r requirements.txt
uvicorn app:app --host 0.0.0.0 --port 7860

First run downloads Whisper medium (~1.42GB) and the risk classifier (~500MB).

Important: Do NOT use --reload — it restarts the process and re-downloads models.

API Endpoints

Method	Path	Description
GET	`/`	Status check
GET	`/health`	Health check (container orchestration)
GET	`/docs`	Swagger UI
POST	`/analyze_audio`	Main endpoint — analyze audio file

Tech Stack

Component	Technology
Language detection	OpenAI Whisper (medium, audio-based)
Speech-to-text	Google Speech Recognition API
Translation	deep-translator (Google Translate)
Romanization	Google Translate romanization API
Risk classification	HuggingFace Transformers (sohampal0011/risk-classifier)
Web framework	FastAPI + Uvicorn
Runtime	Python 3.10, PyTorch CPU
Deployment	Docker on HuggingFace Spaces

Supported Languages

Language	Detection	Transcription	Script
Hindi	✅ Whisper	✅ Google STT	Devanagari
Bengali	✅ Whisper	✅ Google STT	Bengali
English	✅ Whisper	✅ Google STT	Latin
18+ others	✅ Whisper	✅ Google STT	Various

Audio Requirements

Constraint	Value
Format	WAV only
Max duration	35 seconds
Sample rate	Any (48kHz recommended)
Channels	Mono or stereo

Performance

Metric	Value
Cold start	~20-30s
Warm request (English)	~3-5s
Warm request (Hindi/Bengali)	~5-8s
Memory	~1.5-2 GB

Documentation

See API_REQUIREMENTS.md for detailed technical documentation including:

—Full architecture diagrams
—Build pipeline details
—Deployment challenges & solutions
—Integration guide with code examples

License

MIT