> ## 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.

# Datasets

> Every public dataset behind Modulate's published benchmark results, the attack family or acoustic condition each one stresses, and the conditions none of them cover.

*Reviewed as of 19 August 2026.*

Modulate's published results are measured on the datasets its benchmarks define. Neither corpus set is chosen by Modulate.

## Speech DF Arena: 14 deepfake datasets

Speech DF Arena scores deepfake detection across 14 evaluation sets spanning synthesis families, languages, and channel conditions.

| Dataset | Attack family | Condition it stresses |
| - | - | - |
| ASVspoof 2019 | Text-to-speech and voice conversion | Studio-clean synthesis. The long-standing baseline, and the set most detectors are tuned against. |
| ASVspoof 2021 LA | Text-to-speech and voice conversion | The 2019 attacks after transmission through telephony codecs. Separates codec robustness from detection ability. |
| ASVspoof 2021 DF | Compressed synthetic speech | Varied lossy encoders and bitrates, as media re-encoding would apply. |
| ASVspoof 2024 Eval | Crowdsourced and adversarial synthesis | Newer generation methods over non-studio source recordings. |
| Fake or Real | Commercial text-to-speech | Synthesis from deployed commercial systems, across mixed recording conditions. |
| Codecfake | Neural audio codec resynthesis | Speech reconstructed through a learned codec rather than a vocoder. Genuine speech enters the pipeline, so there are no synthesis artefacts to key on. |
| ADD 2022 Track 1 | Mandarin full-utterance fakes | Low-quality and noisy Mandarin audio. |
| ADD 2022 Track 3 | Adversarial Mandarin fakes | Attacks constructed to evade detection rather than to sound natural. |
| ADD 2023 Round 1 | Mandarin deepfake detection | Second-edition challenge audio, first evaluation round. |
| ADD 2023 Round 2 | Mandarin deepfake detection | Second evaluation round of the same challenge. |
| DFADD | Diffusion and flow-matching text-to-speech | Generation families that postdate most detectors' training data. |
| LibriVoc | Vocoder artefacts | Multiple vocoders over read speech, isolating the vocoder as the only synthetic signal. |
| SONAR | Recent end-to-end text-to-speech | Current-generation systems, including synthesis with no separate vocoder stage. |
| In The Wild | Real-world deepfakes | Deepfaked speech of public figures collected from social media. No controlled synthesis pipeline and no matching training distribution. |

In The Wild is the only set drawn from deepfakes made to deceive rather than to populate a corpus. Codecfake is the only set whose audio starts as genuine speech, which defeats detectors keyed to synthesis artefacts.

Results: [Deepfake Detection](/benchmarks/deepfake-detection).

## Open ASR Leaderboard: 8 transcription datasets

The Open ASR Leaderboard scores English transcription across eight corpora, applying one text normaliser to every entrant.

| Dataset | Speech type | Condition it stresses |
| - | - | - |
| AMI | Multi-party meetings | Far-field microphones, overlapping speech, spontaneous turn-taking. The hardest set on the leaderboard for every entrant. |
| Earnings-22 | Earnings calls | Telephony-grade audio, accented English, dense financial vocabulary and named entities. |
| GigaSpeech | Mixed podcast, audiobook, and video | Broad-domain spontaneous speech across recording qualities. |
| LibriSpeech clean | Read audiobooks | Well-recorded read speech. Near-saturated across entrants. |
| LibriSpeech other | Read audiobooks | The harder speaker split of the same corpus. |
| SPGISpeech | Financial calls | Long-form professional speech with heavy domain vocabulary. |
| TED-LIUM | Conference talks | Prepared single-speaker delivery, varied accents. |
| VoxPopuli | European Parliament proceedings | Non-native English accents and parliamentary register. |

Earnings-22 and VoxPopuli are the two sets closest to contact-centre audio. Modulate reports that pair as a named subset alongside the full leaderboard result.

Results: [Transcription](/benchmarks/transcription).

## What these datasets do not cover

* **8 kHz telephony throughout.** ASVspoof 2021 LA applies telephony codecs and Earnings-22 is telephony-grade, but most sets in both benchmarks are wideband. A pipeline running narrowband PCM end to end is outside the measured distribution.
* **Non-speech audio between speech.** IVR prompts, hold music, ringback, and transfer tones are absent from every set listed here.
* **Streaming.** Both benchmarks score complete files. Neither measures partial-result latency or accuracy on a live socket.
* **Languages beyond the sets listed.** Speech DF Arena covers English and Mandarin. The Open ASR Leaderboard is English only. Neither speaks to the rest of the languages [Multilingual Transcription](/get-started/stt) accepts.
* **Speaker diarization, emotion, accent, and PII/PHI.** No public benchmark in this tab scores them. Modulate evaluates them internally.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.