Audio Technical Integrity Check

AI analysis of spectral characteristics and prosodic patterns to detect audio manipulation.

The verify-audio-technical pipeline assesses whether an audio recording is technically consistent with authentic, unmodified speech or sound. It analyses the frequency spectrum for splice artefacts and synthesis signatures, and evaluates prosodic patterns — rhythm, pitch, and stress — for characteristics associated with synthetic or edited audio. This pipeline is well suited to triage workflows, newsroom intake, and any situation where you need a fast automated signal on a spoken-word recording before investing in a full crowd-reviewed deepfake analysis.

The pipeline does not attempt to identify the speaker or verify the content of what is said. Its focus is on the technical integrity of the recording itself: whether it is a single, continuous, authentic capture or whether it shows evidence of having been assembled, altered, or generated.

Pipeline Details

Property	Value
Pipeline ID	`verify-audio-technical`
Tier	1 — AI-Only
Estimated Duration	~1–3 min
Credit Cost	900 credits per run
Crowd Stages	No
Files Required	Yes — enriched audio (MP3, WAV, FLAC, OGG, or M4A)

Required Inputs

Files

One enriched audio file. The file must have completed enrichment before the run is created (enrichmentStatus: "completed"). See Files & Enrichment for details.

Context Keys

Both keys are optional but help the AI contextualise its analysis — particularly the prosody evaluation, which is more meaningful when the AI knows what kind of speaker is being assessed.

Key	Required	Description
`claimed_speaker`	Optional	Name or role of the claimed speaker (e.g. `"Press Secretary Jane Doe"`, `"Senator John Smith"`). Used to contextualise prosodic expectations.
`claimed_date`	Optional	Claimed recording date in ISO 8601 format (e.g. `"2024-10-03"`). Cross-referenced against encoding metadata where available.

Stages

Spectral Analysis (AI)

Analyses the full frequency spectrum of the audio file for technical artefacts associated with manipulation. The AI looks for discontinuities in the spectral envelope — abrupt changes in noise floor, frequency cutoffs that appear mid-file, or phasic inconsistencies that indicate content from different recordings has been joined. It also checks for patterns characteristic of audio synthesis, such as the unnaturally flat frequency distribution produced by some text-to-speech systems or the characteristic formant structures of voice conversion models.

Prosody Analysis (AI)

Evaluates the rhythm, pitch contour, and stress patterns of any speech present in the recording. Human speech has natural micro-variations in timing and intonation that are difficult to replicate perfectly with synthesis. The AI assesses whether the prosodic patterns are consistent with spontaneous authentic speech, read speech, or machine-generated speech. Unusual flatness in pitch variance, overly regular inter-word gaps, or stress patterns that do not match the semantic content of the utterance are all flagged as signals. This stage writes a syntheticLikelihood score (0–100) and a list of syntheticIndicators to its artifacts, and sets its verdict to synthetic when TTS/voice-clone evidence is strong — see Synthetic Data Detection.

Verdict Synthesis (AI)

Aggregates the spectral and prosodic signals into a unified final verdict. Each signal is weighted by confidence — strong spectral splice evidence, for instance, carries more weight than mild prosodic irregularity — and the AI produces a verdict with a confidence score and an explanation of the most significant contributing factors.

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-technical",
    "file_ids": ["{fileId}"],
    "context": {
      "claimed_speaker": "Press Secretary Jane Doe",
      "claimed_date": "2024-10-03"
    }
  }'

Example Response

{
  "id": "run_01j9x...",
  "status": "completed",
  "verdict": "authentic",
  "confidence": 82,
  "pipeline_id": "verify-audio-technical",
  "scorecard": {
    "spectral_continuity": "pass",
    "noise_floor_consistency": "pass",
    "prosody_naturalness": "pass"
  },
  "explanation": "Spectral analysis shows a consistent noise floor with no detectable splice points. Prosodic patterns are consistent with authentic spontaneous speech. No synthesis signatures detected."
}

For deepfake detection with native-speaker crowd review and audio-expert sign-off — which significantly increases detection reliability for high-quality voice clones — see verify-audio-deepfake.