> ## Documentation Index
> Fetch the complete documentation index at: https://docs.modulate.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Deepfake Detection results

> Speech DF Arena results for Modulate's Deepfake Detection model: pooled and average EER across 14 datasets, competitor scores, and deployment characteristics.

*Reviewed as of 19 August 2026.*

[Deepfake Detection](/get-started/deepfake) is ranked 1st on [Speech DF Arena](https://huggingface.co/spaces/Speech-Arena-2025/Speech-DF-Arena), with an average EER of 1.104% across the arena's [14 evaluation datasets](/benchmarks/datasets#speech-df-arena-14-deepfake-datasets).

<Note>
  **Snapshot: 19 August 2026.** Speech DF Arena accepts submissions continuously, so this standing is accurate on that date and not after. The [live leaderboard](https://huggingface.co/spaces/Speech-Arena-2025/Speech-DF-Arena) is authoritative. Modulate's row is listed under a legacy `VELMA-2` identifier rather than the canonical model name.
</Note>

## Scores

Top four systems at the snapshot date. Lower is better in both EER columns.

| System                          | Average EER | Pooled EER | Detection accuracy |
| ------------------------------- | ----------- | ---------- | ------------------ |
| **Modulate Deepfake Detection** | **1.104%**  | **1.1%**   | **98.9%**          |
| Hiya Authenticity Verification  | 2.113%      | 2.324%     | 97.9%              |
| Resemble Detect 3B Omni         | 2.570%      | 2.099%     | 97.4%              |
| Whispeak                        | 3.05%       | 3.00%      | 96.9%              |

Detection accuracy is `100% - average EER`. It restates the first column.

The two EER columns order these systems differently. Resemble Detect 3B Omni posts the second-best pooled EER and the third-best average EER; Hiya is the reverse. Deepfake Detection leads on both, so one threshold holds across all 14 datasets. [Pooled EER and average EER](/benchmarks/methodology#pooled-eer-and-average-eer) covers the distinction.

### At an equal-error operating point

At the equal-error threshold the same rate applies in both error directions. Each figure below is simultaneously the missed synthetic calls per 1,000 synthetic calls and the false alarms per 1,000 genuine calls.

| System                          | Errors per 1,000 calls |
| ------------------------------- | ---------------------- |
| **Modulate Deepfake Detection** | **\~11**               |
| Hiya Authenticity Verification  | \~21                   |
| Resemble Detect 3B Omni         | \~26                   |
| Whispeak                        | \~31                   |

This is arithmetic on the average EER column, not a separate measurement. The next-best system by average EER produces roughly 1.9 times as many errors at this operating point.

<Warning>
  **Worth knowing:** production systems rarely run at the equal-error threshold. Moving the threshold to catch more synthetic calls raises false alarms on genuine ones, and the published EER does not describe that curve. Set the threshold against your own labelled recordings, as [voice fraud screening](/get-started/voice-fraud-screening) describes.
</Warning>

## Deployment characteristics

| Property                    | Modulate Deepfake Detection            | Leaderboard competitors |
| --------------------------- | -------------------------------------- | ----------------------- |
| Parameter count             | 316 million                            | Over 1 billion          |
| Minimum audio for a verdict | 2.5 seconds                            | 5 to 30 seconds         |
| Published price             | \$0.25 per hour                        | $29 to $120 per hour    |
| Streaming support           | Yes, per-frame verdicts over WebSocket | Varies                  |

Minimum audio duration determines whether a model can gate a live authentication flow. Deepfake Detection returns a verdict per 192 ms frame once 2.5 seconds of speech has accumulated.

Competitor prices are list prices at the snapshot date. The arena does not measure them.

## What the arena does not measure

* **Streaming behaviour.** The arena scores complete files. The per-frame latency of [Deepfake Detection streaming](/api-reference/svd/streaming) is not part of this result.
* **Narrowband telephony end to end.** ASVspoof 2021 LA applies telephony codecs. Most arena datasets are wideband.
* **Non-speech audio.** Hold music, IVR prompts, and ringback are absent from every arena dataset. In production those frames return `verdict: "no-content"` and are excluded from scoring.
* **Languages beyond English and Mandarin.**

## Sources

* [Speech DF Arena leaderboard](https://huggingface.co/spaces/Speech-Arena-2025/Speech-DF-Arena), Hugging Face. Live table.
* [Speech DF Arena: A Leaderboard for Speech DeepFake Detection Models](https://arxiv.org/abs/2509.02859), arXiv:2509.02859. Benchmark design, dataset selection, metric definitions.
* [speech\_df\_arena toolkit](https://github.com/Speech-Arena/speech_df_arena), GitHub. Evaluation harness and protocol format.