Velma Deepfake Detect
AI model · Safety / moderation · Text → Tables / structured data by Modulate
It identifies AI-generated, cloned, or synthetic voices in both recorded files and live audio streams with a 98.9% accuracy rate. The model is built on an Ensemble Listening Model (ELM) architecture, allowing it to process audio natively and maintain high performance in noisy environments. It is delivered via REST and WebSocket APIs, making it suitable for integration into contact centers, fintech platforms, and authentication flows.
- Languages
- English
- Input Modality
- Audio (File or Stream)
- Output Modality
- Synthetic-voice verdict (Probability)
- Parameter Count
- 316 Million
- Equal Error Rate (EER)
- 1.1%
- Minimum Audio Required
- 2.5 seconds
- Access Method
- REST API / WebSocket API
Overview
It identifies AI-generated, cloned, or synthetic voices in both recorded files and live audio streams with a 98.9% accuracy rate. The model is built on an Ensemble Listening Model (ELM) architecture, allowing it to process audio natively and maintain high performance in noisy environments. It is delivered via REST and WebSocket APIs, making it suitable for integration into contact centers, fintech platforms, and authentication flows.
Detection Capabilities
- 98.9% Accuracy: The model identifies AI-generated, cloned, or synthetic voices with 98.9% accuracy.
- Minimal Audio Requirement: The model requires only 2.5 to 3 seconds of audio to produce a reliable verdict.
- Continuous Monitoring: Velma continuously monitors the whole call, detecting synthetic voice switches that occur mid-conversation.
Technical Performance
- Noise Resilience: The model is optimized for noise resilience and remains accurate in messy, real-world audio environments.
- Cost Efficiency: Velma operates at a 120x lower cost than traditional deepfake detection solutions.
- Flexible Deployment: The model is available as both a batch REST API for file analysis and a real-time streaming API over WebSocket.
Pricing
- Batch (per hour)
- $0.25
- Streaming (per hour)
- $0.25
- Minimum per API call
- $0.05
Specifications
- Input Modality
- Audio (File or Stream)
- Output Modality
- Synthetic-voice verdict (Probability)
- Parameter Count
- 316 Million
- Equal Error Rate (EER)
- 1.1%
- Minimum Audio Required
- 2.5 seconds
- Access Method
- REST API / WebSocket API
- Hugging Face Rank
- #1 on Speech Arena Leaderboard
Added by AntSpace from public sources (modulate.ai).