Description
VeriSpoof is a pioneering audio language model developed for advanced speech deepfake forensics. It provides a comprehensive solution for determining whether speech is a deepfake, identifying the system that generated it, assessing if two recordings share a common source, analyzing channel effects, and explaining its conclusions in a transparent, verifiable manner. This single, unified audio LLM (Large Language Model) operates beyond simple real or fake classifications, offering detailed insights into the origin and manipulation of audio.
Read More
The model has demonstrated exceptional performance, achieving a 2.14% pooled Equal Error Rate (EER) across 14 Speech-DF-Arena sets, positioning it as the top-ranked open-weight detector. Its architecture is built around an audio LLM that processes recordings into utterance and frame tokens, which are then read by a MiniCPM5-2B model with LoRA. Unlike traditional systems, VeriSpoof does not rely on separate task heads; instead, every answer is a sentence whose likelihood directly serves as the score, allowing for flexible thresholding at any operating point.
VeriSpoof's grounding is a core strength, with each line of its rationale independently checked against waveform measurements, logged conditions, labeled spans, and generator metadata. This ensures that observations, channel status, localization of synthesized segments, and source profiles are rigorously validated. The system also incorporates counterfactual training, where decisions are trained to follow stated evidence, even when localization is intentionally reversed, ensuring robust and logically consistent reasoning.
Intended primarily for research in deepfake detection, source tracing, and explainable forensics, VeriSpoof provides a powerful tool for auditing audio verdicts. It is recommended that its scores be combined with other evidence and human review for final forensic decisions. The project includes both a Base and a Plus model, with VeriSpoof-Plus demonstrating superior performance across various evaluation metrics and datasets. The weights and code for VeriSpoof are released under a CC BY-NC 4.0 license, while its base models, MiniCPM5-2B and Qwen3-ASR-1.7B, operate under Apache-2.0.
- CategoryAI Content Detectors
- Pricing ModelFree (Open-Weight)
- DeploymentLocal (Inference Package)
- PlatformsWindows, macOS, Linux
VeriSpoof's Core Features
Speech Deepfake Detection: Identifies whether speech is a deepfake.
Source Verification: Determines which system generated the deepfake.
Source Attribution: Ascertains if two recordings originate from the same generator.
Channel Description: Analyzes and describes the channel effects on audio.
Explainable Decisions: Provides short, verifiable lines explaining its conclusions.
Grounded Rationale: Each claim is checked against waveform measurements and dataset labels.
High Accuracy: Achieved 2.14% pooled EER on Speech-DF-Arena, ranking #1 among open-weight detectors.
Unified Audio LLM: One model handles all forensic questions without separate task heads.
Score-Based Answers: Answers are sentences whose likelihood serves as the score.
Counterfactual Testing: Allows replacing stated localization to re-evaluate decisions.
Offline Inference: Supports running real test recordings offline.
Open-Weight Model: Available for research and development.
Multilingual Captions: Film demonstrations include Chinese and English captions.
Research-Oriented: Designed for deepfake detection, source tracing, and explainable forensics research.
Auditable Verdicts: Four lines to a verdict that can be audited.
How to use VeriSpoof?
Input Audio: Provide a single audio recording or a pair of recordings to the model.
Process with LLM: The audio components convert the input into utterance and frame tokens for the LLM.
Receive One-Step Answer: Obtain an immediate answer regarding deepfake status or source similarity.
Review Grounded Rationale: Examine the detailed, line-by-line explanation for the model's decision.
Verify Claims: Check each claim in the rationale against waveform measurements and dataset labels.
Conduct Counterfactual Tests: Optionally, modify parameters like localization and re-run the model to observe decision changes.
VeriSpoof's Use Cases
- Deepfake Detection Research
- Audio Source Tracing
- Explainable Forensics
- Audio Channel Analysis
- Synthetic Speech Analysis
- Benchmarking Deepfake Detectors








