DotCheck/helmholtz-audio-v3

Helmholtz is DotCheck’s production audio detector. We train a logistic head on frozen Dasheng-Base embeddings, gate on held-out speech, music/SFX, and fal-family TTS, and publish the holdout table below.

Open package: live .npz head, card, license, notices. In product, soundtrack windows can fuse with Muybridge frame bags (Covenant); that seating stays private. This repo is standalone audio.

Field	Value
Hub id	DotCheck/helmholtz-audio-v3
Wire id	inhouse-audio@3
Label	Helmholtz
Artifact	dasheng_base_linear_head_audio_v3.npz
Backbone	mispeech/dasheng-base (Apache-2.0)
Protocol	4 s / center crop / up to 2 windows / max aggregate (serve may single-pass prewindowed clips)
Output	p ∈ [0,1] — estimated P(AI)
Serve	CPU FastAPI /v1/analyze-audio; Express Check + video soundtrack fuse

Model description

Frozen Dasheng-Base masked audio encoder + logistic head for speech, music, and general audio. Product Covenant seats soundtrack windows with Muybridge frame bags when a video has usable audio. Covenant is DotCheck-private; these files are the open audio head only.

In this repo: README.md, LICENSE, NOTICE, CITATION.cff, and the .npz head.

Architecture

audio bytes
  → mono 16 kHz
  → protocol window(s) (4 s)
  → frozen Dasheng-Base emb → logistic head → p
  → multi-window: max

Missing audio head at serve → /v1/analyze-audio returns 503 (fail closed).

Inference

Open weights: the live .npz head (Apache-2.0) with the frozen backbone above. Not AutoModel.from_pretrained("DotCheck/…").

Product scoring: Check or Pro API.

curl -sS -X POST "https://dotcheck-server-c221c1f32c68.herokuapp.com/analyze-audio" \
  -H "Authorization: Bearer dc_YOUR_KEY" \
  -F "audio=@clip.wav"

Gates: dotcheck.ai/docs.

Training data

Split	Content
Fit reals	LibriTTS + FSD50K (concat, no loop-tile)
Fit AI	parler-mini / melo-en / dasheng-audiogen + fal train (Kokoro / Eleven Turbo / mixes) + 22 Imagine few-shot
Holdout speech	CodecFake / DFADD (sha-disjoint)
Holdout music/SFX	AudioGen odd-seed partition
Holdout fal	prompt-id holdout (same gens, disjoint sha)

Evidence: audio_gates_audio_v3.json · AUDIO_GATES_OK. Imagine-owned audio is a watch, not a public claim.

Evaluation

Holdout (public claim SSOT)

Metric	Target	Measured
mean P(AI) | real	≤ 0.12	0.005
mean P(AI) | AI	≥ 0.85	0.982
bal_acc @ thr	≥ 0.90	0.989

Intended use

• Standalone audio AI detection in DotCheck inference.
• Soundtrack windows fused with video frames in product serve.
• Citation of the holdout table above.

Out of scope

• Product SLA via Hub download alone
• Courtroom deepfake forensics or speaker identification
• Legal determinations

Limitations

• Protocol windows are short; long-form structure is not modeled.
• Codec, compression, and unseen generators can shift scores.
• Public floats are standalone audio gates; fused video bags are not a separate claim table here.

License

LICENSE — Apache License 2.0 for DotCheck heads in this repo. Upstream: NOTICE.

Citation

CITATION.cff · wire inhouse-audio@3 / Helmholtz@3 · https://dotcheck.ai/docs