Audio Deepfake Detection

AI spectral and prosodic analysis combined with native-speaker crowd review to detect synthetic or manipulated audio.

The Audio Deepfake Detection pipeline combines low-level AI signal analysis with human native-speaker review to identify synthetic speech, voice cloning, and audio splicing. While AI models are effective at detecting the technical fingerprints left by common voice synthesis and conversion systems, they have well-documented blind spots around prosodic naturalness, regional accent consistency, and emotional register coherence. Native-speaker crowd workers fill precisely this gap: a fluent listener can detect the subtle flatness of a cloned voice or the wrong intonation pattern for a claimed speaker in ways that spectral analysis cannot reliably capture. This pipeline is well suited for fact-checkers evaluating contested audio recordings, legal teams assessing voice evidence, and newsrooms deciding whether to publish audio-based reporting.

Pipeline Details

Property	Value
Pipeline ID	`verify-audio-deepfake`
Tier	2
Estimated Duration	~60 min
Credit Cost	2,900 credits per run
Crowd Stages	Yes
Files Required	Yes — enriched audio

Required Inputs

Files

Submit one enriched audio file. Enrichment must be complete before starting this pipeline. Accepted formats: MP3, WAV, FLAC, AAC, OGG, M4A.

Context Keys

Key	Required	Description
`claimed_speaker`	Optional	Identity of the person claimed to be speaking (e.g. `"Chancellor Schmidt"`)
`claimed_language`	Optional	BCP-47 language tag for the audio's spoken language (e.g. `"de"`, `"en-US"`, `"fr"`); used to route the task to appropriately fluent workers
`claimed_date`	Optional	Date the recording is claimed to have been made, ISO 8601 format
`claimed_event`	Optional	Event or context for the audio recording (e.g. `"Press conference 2024-11-20"`)

Setting claimed_language is strongly recommended. It allows the platform to route the crowd stage to workers who are fluent in the spoken language, which is essential for meaningful prosodic and accent assessment. Without it, routing defaults to English-fluent workers.

Stages

Technical Forensics (AI)

The pipeline begins with a forensic analysis of the audio file's technical properties. The AI examines spectral features across frequency bands, compression history (double-encoded segments, re-quantisation artefacts), encoding parameter consistency, and DC offset irregularities. Segments that show statistical profiles inconsistent with natural recording conditions are flagged with timestamps and confidence scores. These findings form the technical foundation for the stages that follow.

Deepfake Indicators (AI)

The AI applies synthesis-specific detection models to the full audio. This stage evaluates vocoder fingerprints left by common text-to-speech and voice-conversion systems, prosodic unnaturalness (pitch contour regularity, abnormal pause distribution, missing co-articulation), and acoustic inconsistencies between segments that may indicate splicing. The output is a per-segment anomaly timeline and an aggregate synthesis probability score covering the full recording.

Native Speaker Review (Crowd)

Three crowd workers holding the native_speaker skill tag — matched to the audio's claimed or detected language — listen to the full recording and complete a structured assessment questionnaire. Workers evaluate overall naturalness, speaker consistency across segments, accent plausibility relative to any claimed identity, emotional register coherence, and whether any segments sound computer-generated or edited. The stage gates on 70% consensus (at least two of three workers in agreement) before proceeding. Workers are shown the AI technical findings as context but are asked to form their perceptual judgment independently before reviewing them.

Verdict Synthesis (AI)

The AI integrates the technical forensic findings, the deepfake indicator scores, and the structured crowd worker assessments into a final weighted verdict. Divergent worker responses are explicitly noted in the synthesis output. The final result includes a confidence-calibrated verdict (authentic, manipulated, or inconclusive), an anomaly summary with timestamped references to flagged segments, and a complete per-stage evidence log.

Starting a Run

curl -X POST "https://api.crowdee.ai/v2/projects/{projectId}/verification-runs" \
  -H "X-API-Key: crw_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "pipeline_id": "verify-audio-deepfake",
    "file_ids": ["file_01j9x..."],
    "context": {
      "claimed_speaker": "Chancellor Schmidt",
      "claimed_language": "de",
      "claimed_event": "Press conference 2024-11-20"
    }
  }'

Example Response

{
  "id": "run_01j9x...",
  "status": "completed",
  "verdict": "authentic",
  "confidence": 81,
  "pipeline_id": "verify-audio-deepfake"
}

For faster AI-only audio analysis, see verify-audio-technical. Setting claimed_language helps route the crowd stage to workers fluent in the language of the recording — this meaningfully improves the quality of the native-speaker review.