Modelpublic

DotCheck/helmholtz-audio-v3_1

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes

Model Card

DotCheck/helmholtz-audio-v3_1

`Helmholtz@3.1` (inhouse-audio@3) is an audio detector. It maps a waveform to \(p \in [0,1]\), an estimate of \(P(\mathrm{AI})\) for speech, music, and general sound.

The product is Helmholtz: a logistic head we trained on frozen Dasheng-Base embeddings (the ancestor encoder, parallel to SigLIP 2 on the vision cards). It is not a fine-tuned encoder. In the DotCheck product, soundtrack windows on a video can be fused with Muybridge frame scores (Covenant). That fusion is a private assembly rule and is not a public claim in this repository. The table below is standalone audio.

Field	Value
Hub id	`DotCheck/helmholtz-audio-v3_1`
Wire id	`inhouse-audio@3`
Public token	`Helmholtz@3.1`
Artifact	`dasheng_base_linear_head_audio_h3e.npz`
Backbone	`mispeech/dasheng-base` (Apache-2.0)
Window protocol	4 s, center crop, up to 2 windows, aggregate = max
Output	\(p \in [0,1]\) — estimated \(P(\mathrm{AI})\)
Serve	CPU FastAPI `/v1/analyze-audio`

Model description

Input audio is converted to mono 16 kHz. The protocol takes up to two 4-second windows (center crop). Each window is embedded by frozen Dasheng-Base; the logistic head produces \(pw\). Multi-window clips use \(\maxw p_w\). A pre-windowed clip may be scored in a single pass.

In this repo: README.md, `LICENSE`, `NOTICE`, `CITATION.cff`, and the .npz head. The Dasheng checkpoint is not redistributed here.

Architecture

text

audio bytes
  → mono 16 kHz
  → up to two 4 s windows (center crop)
  → frozen Dasheng-Base embedding
  → logistic head (audio_h3e.npz) → p_w
  → clip p = max(p_w)

Inference

Windows are mono 16 kHz, up to two 4 s center crops, max over window scores. Missing head at serve is 503 (fail closed). In product video intake, a RIFF/WAVE window with mono-16-bit RMS \(\le 10^{-4}\) is omitted before this head (silence is not scored as \(p=0\)).

Open weights: the live .npz head in this repository (Apache-2.0), used with the frozen Dasheng-Base backbone named above. This is not a transformers AutoModel package.

Product scoring: Check or Pro API. Hub storefront: `see-whats-real`.

Training data

Split	Content
Fit reals	LibriTTS + FSD50K (concatenated; no loop-tile)
Fit AI	parler-mini / melo-en / dasheng-audiogen + fal train (Kokoro / Eleven Turbo / mixes) + Imagine in the linear mix
Holdout speech	CodecFake / DFADD (SHA-disjoint)
Holdout music/SFX	AudioGen odd-seed partition
Holdout fal	prompt-id holdout (same generators, disjoint SHA)
Holdout Imagine	owned exam (Layer A); not a separate public row

Evidence: audio_gates_audio_h3e.json (AUDIO_GATES_OK). Public floats are the overall holdout (claims.audio in Data.json).

Evaluation

Binary classification at threshold \(0.5\). Public claim = overall holdout class-conditional mean \(P(\mathrm{AI})\) and balanced accuracy.

Metric	Floor	Measured
mean \(P(\mathrm{AI})\) \	real	\(\le 0.12\)	0.018
mean \(P(\mathrm{AI})\) \	AI	\(\ge 0.85\)	0.989
balanced accuracy	\(\ge 0.90\)	0.995

Intended use

—Reproduce the logistic head and the holdout table.
—Research on synthetic speech, music, and general-audio detection under this window protocol.

Out of scope: speaker identification, legal determinations, and fused frame+audio scores. The public table is standalone audio.

Limitations

—At most two 4 s windows: long-form structure, sparse events, and late-onset synthesis are not modeled.
—Holdout is CodecFake/DFADD speech, AudioGen music/SFX, and a fal-family SHA split. Unseen codecs and TTS families can shift scores.
—The public floats are overall holdout means at threshold 0.5, not per-domain rows and not a fused video claim.

License

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

Citation

`CITATION.cff` · wire inhouse-audio@3 / Helmholtz@3.1 · https://dotcheck.ai/docs