How the current detector computes acoustic features and an uncalibrated decision score, where its evidence is limited, and which integrations remain design proposals.
1. Acoustic Physics: Human Larynx vs Neural Vocoders
01
Human phonation involves airflow, vocal-fold vibration, and vocal-tract filtering. The detector estimates properties of the recorded signal that may reflect aspects of this process. Those estimates are indirect: microphones, channels, codecs, speakers, and synthesis systems all affect the measured waveform.
Sk=μ3σ3=1N∑(e[i]−e)3(1N∑(e[i]−e)2)3/2
LPC Residual Skewness: The third standardized moment of an estimated residual e[i]. The API field is glottal_skewness. It is an acoustic estimate, not a measurement of vocal-fold motion. The selected benchmark does not establish a population range separating human and synthetic speech.
1. Signed statistic and polarity: Skewness is a signed third standardized moment (S_k = μ_3 / σ^3). A polarity inversion can change its sign, so the detector uses |S_k| in some rules. This handling does not make the feature invariant to codecs or recording conditions.
2. Hand-set operational rules: The JavaScript engine uses |S_k| ≥ 0.45, or |S_k| ≥ 0.38 together with a diagonal coupling estimate of at least 0.25, as a reference-pattern condition. The 0.38 value is not a standalone human-speech floor. These manually chosen constants have not been independently validated as physiological boundaries or calibrated probabilities.
3. Engine-specific rules: The JavaScript API combines skewness with several spectral and reflection features. The Python engine includes additional QPC and phase-entropy score terms, so the engines can disagree. No rule guarantees protection for breathy or non-modal human speech; that requires independent evaluation.
What the numerical rules mean
Research on glottal-flow asymmetry motivates examining waveform shape. That does not equate a flow measurement, clinical waveform descriptor, and this LPC residual's amplitude moment. Feature motivation, measured corpus distributions, and hand-set scoring rules are separate evidence.
JavaScript score component
Effect in the current rules
Interpretation limit
|S_k| ≥ 0.45, or |S_k| ≥ 0.38 && coupling ≥ 0.25
Enables a reference-pattern condition; further spectral and reflection checks determine whether it lowers the score.
This is neither a universal biological range nor an independent classification.
|S_k| < 0.18
Adds 3.00 to the combined pre-logistic score.
Low residual skew can occur for reasons other than synthesis.
jitter < 0.35%
Adds 3.20 to the combined pre-logistic score.
This percentage concerns estimated period variation; it is not a skewness cutoff or a physiological minimum.
decision_score ≥ τ
Flags synthetic speech after applicable speech checks and combined scoring. The default is τ = 0.50.
Raising the threshold flags fewer files. The score is uncalibrated; its value is not the probability that a verdict is correct.
These are selected components, not the whole decision function. The release manifest identifies the exact public DSP source. The 29-file and 168-trial Worker ledgers were scored with that source. Each older research study (ODSS, ASVspoof 5, and the SSL comparison) names the earlier frozen source it used, and the current source was not run on those studies. The measured evaluations show how the combined rules fail under domain shift.
Diagonal bispectral coupling estimate
The API field qpc_bicoherence samples only diagonal frequency triads (k, k, 2k). For the processed frames m, the implementation computes B(k) = mean_m[X_m(k)^2 conjugate(X_m(2k))] and P(k) = mean_m[|X_m(k)|^2]. It averages min(1, |B(k)|^2 / (P(k)^2 P(2k) + 10^-12)) over sampled bins.
The JavaScript routine uses 256-point Hann-windowed frames, a nominal 128-sample hop, every other frame, and every other selected frequency bin, on 16 kHz analysis audio. It does not compute a full two-dimensional bicoherence map. This particular normalization, subsampling, and clipping define the feature; its value should not be treated as a direct measure of tissue collision or a universal phonation threshold.
Speech synthesis systems use different architectures. Some vocoders, including HiFi-GAN, use transposed-convolution upsampling; others, such as Vocos, synthesize spectral coefficients and apply an inverse Fourier transform. A feature hypothesis should be tested against each architecture and recording channel.
The detector explores several possible signal differences:
Phase structure: Some synthesis and upsampling methods can leave periodic artifacts. A phase measurement alone does not prove which method produced a recording.
Pitch variation: Jitter features may vary across speakers and generators, as well as with tracking errors and channel effects. No universal synthetic cutoff has been established.
Spectral residuals: Flow-matching and diffusion outputs may have measurable residual patterns, but an ODE or diffusion mechanism cannot be inferred from a single high-frequency value.
2. Patent Context for Audio Buffering
US Patent 11,924,717, System and method for data analytics for communications in a walkie-talkie network. Its claims cover walkie-talkie communication through a mobile device and remote processing; some claims mention a circular buffer.
The patent is background for packet and buffer design. It does not establish the synthetic-voice detector's feature validity, a SIP/RTP implementation, jitter-buffer compensation, or any measured latency. This repository does not include evidence mapping a specific patent claim to the current detector. Acoustic scoring and receipt signing are separate project work.
3. Current File Verification Pipeline
03
The four stages below describe the current JavaScript API engine conceptually. A request supplies a complete supported PCM recording; the service does not accept RTP packets or emit per-frame verdicts. No stage-level production latency distribution has been measured.
1. Decode and screen
Validate the WAV or raw PCM input, normalize the audio used for analysis, and identify voiced content. Silence and non-vocal inputs can return an inconclusive verdict.
2. Extract acoustic features
Compute phase, spectrum, pitch, residual, jitter, and reflection statistics. These are signal estimates affected by recording conditions; they do not directly observe vocal-fold anatomy or identify a generator.
3. Apply decision rules
Combine hand-set feature terms into a sigmoid-shaped detection score and compare it with the requested threshold. The canonical decision_score is uncalibrated; it is not a posterior probability.
4. Return result and optional receipt
Return a verdict, score, and optional telemetry. If a signing key is configured, the service can attach an Ed25519 vendor-format receipt for payload integrity. Signature verification alone does not establish issuer trust or empirical correctness.
Python, JavaScript API, and TypeScript local engines can produce different scores on the same audio. The cross-engine benchmark reports those differences; client code should identify which engine produced a result.
3.5 Browser Visualization and SDK Chunks
CURRENT BEHAVIOR
The browser demo can visualize incoming audio locally. SDK chunk helpers submit each supplied chunk as an independent HTTP verification request. They do not provide a server-side audio session, an utterance latch across requests, a per-frame alert guarantee, or measured submillisecond end-to-end latency. Short or unvoiced chunks may be inconclusive.
A stateful streaming product would need codec adapters, buffered utterance policy, sequence handling, authentication, backpressure, and measured latency under concurrent load. That work remains a design proposal.
4. Codec & Channel Evaluation
04
The current REST endpoint accepts supported PCM WAV encodings and raw 16 kHz PCM16. It does not decode G.711, Opus, AAC, or MP3 payloads directly. A caller must decode those formats into supported PCM before verification.
Selected derived samples
The local stress corpus includes 18 WAVs made from six source clips with G.711 μ-law and two Opus settings. The current JavaScript engine classified all 18 at threshold 0.50. The source clips repeat across codec variants, and the exact encoder commands and versions are not preserved, so this is a regression observation rather than a general codec-robustness estimate.
Open validation work
A deployment study needs disjoint speakers and generators, reproducible codec pipelines, diverse channels and noise, and separate results for each condition. The current benchmark shows poor generalization on its selected In-the-Wild subset; codec claims must be evaluated against that limitation.
The ASVspoof 2021 evaluation plan distinguishes telephony transmission in LA from general compression in DF. Results from one condition should not be presented as proof for another.
5. Threat Model and Evasion Limits
05
An attacker may choose a new generator, post-process audio, replay it through a channel, or adapt samples after observing detector output. Acoustic artifacts are useful features only while they remain associated with the target threat under those changes. No claim is made here that evasion requires a particular compute budget or a new vocal-fold simulator.
Observed generalization gap
On 40 selected In-the-Wild speech trials, the current JavaScript engine reached 37.50% accuracy at threshold 0.50 and 65.00% score-based EER. These convenience samples do not measure population performance, but they rule out a broad robustness claim from this evidence.
Operational safeguards to evaluate
Test representative traffic, track drift by channel and generator, retain a review path for consequential decisions, and validate any threshold before use. Adaptive attacks and benign channel changes both belong in a future red-team protocol.
The Carlini and Wagner speech-to-text attack demonstrates targeted white-box adversarial examples against DeepSpeech. It does not establish how this detector or telephony codecs respond to such attacks.
6. Implemented Interfaces and Proposed Integrations
CURRENT AND DESIGN
Current implementation
A Cloudflare Pages Function verifies complete PCM audio requests. Python and TypeScript SDKs call that API or run a local detector. The main browser demonstration and server scorer use JavaScript. A separate local WebAssembly sandbox runs a compiled C port in a cancellable browser Worker, with shared JavaScript WAV decoding. It does not establish new accuracy or speed claims. SDK chunk helpers make separate HTTP requests; they do not maintain a server-side stream.
Design proposals
A stateful SIP/RTP proxy, WebRTC media adapter, gRPC service, native detector package and server-side WebAssembly deployment remain proposed integration work. None of those services or packages is deployed. The local compiled sandbox is separate from these transport and packaging proposals.
Production latency distributions and end-to-end streaming behavior have not been measured. See the API and SDK reference for supported request formats.
Operational questions
What does the detector threshold mean?
The default threshold is 0.50 on an uncalibrated decision score. A score at or above the threshold flags synthetic speech after applicable speech checks. Raising the threshold flags fewer files. The score is not a probability that the verdict is correct.
Can adaptive attackers evade the detector?
An attacker may change generators, post-process audio, use a different channel, or adapt samples after observing output. No cost or robustness guarantee has been established for this detector. The selected In-the-Wild subset shows a generalization gap.
Does US Patent 11,924,717 validate this detector?
No. US Patent 11,924,717 concerns walkie-talkie communications and remote processing; some claims mention a circular buffer. It does not establish this synthetic-voice detector's feature validity or measured performance.
7. Academic Research Compendium & Citations
07
These sources give context for acoustic synthesis, phonation, spoofing research, and a referenced communications patent. They motivate hypotheses and evaluation design; they do not independently validate this detector's features or performance.
DDSP: Differentiable Digital Signal Processing. International Conference on Learning Representations (ICLR 2020).
Foundational research demonstrating how differentiable classical DSP elements (oscillators, filters, harmonic synthesizers) can be integrated with neural networks, enabling physical acoustic modeling inside backpropagation pipelines.
HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis. Advances in Neural Information Processing Systems (NeurIPS 2020).
Introduces a GAN vocoder with transposed-convolution upsampling and multi-receptive-field fusion. The paper does not validate Noise or Voice's phase metric or a universal artifact frequency; those require separate experiments.
The physics of small-amplitude oscillation of the vocal folds. The Journal of the Acoustical Society of America, 83(4), pp. 1536-1552.
A physical model of small-amplitude vocal-fold oscillation and the conditions for sustained phonation. It does not establish fixed thresholds for deepfake detection.
The Myoelastic Aerodynamic Theory of Phonation. National Center for Voice and Speech. ISBN: 978-0-9772374-1-8.
A monograph on the coupled tissue and airflow mechanisms of phonation. Its physiology should not be treated as a universal distribution for recorded speech features.
Principles of Voice Production. Prentice-Hall. ISBN: 978-0-13-717893-3.
A foundational treatment of voice production, including source-filter concepts and vocal variation. It does not provide a validated cutoff for this detector.
Yamagishi, J., Wang, X., Todisco, M., Sahidullah, M., Patino, J., Nautsch, A., et al.2021
ASVspoof 2021: Accelerating progress in spoofed and deepfake speech detection. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge.
Describes the 2021 challenge tasks, datasets, metrics, baselines, and channel or compression variability. The LA and DF tracks have different conditions.
Audio Adversarial Examples: Targeted Attacks on Speech-to-Text. 2018 IEEE Security and Privacy Workshops (SPW), pp. 1-7.
Demonstrates targeted white-box audio adversarial examples against DeepSpeech. The paper does not test Noise or Voice, its inverse-filter features, or survival through telephony codecs.
WaveFake: A Data Set to Facilitate Audio Deepfake Detection. NeurIPS 2021 Datasets and Benchmarks Track.
Introduces generated-audio samples spanning multiple synthesis architectures and two languages, plus baseline detection models. Its results do not establish a fingerprint for this detector.
System and method for data analytics for communications in a walkie-talkie network. US Patent 11,924,717 B2. Current assignee: VoiceBrain, Inc.; original assignee: VoiceMe.AI, Inc.
Claims walkie-talkie communication through a mobile device and remote computer processing; some claims include a circular buffer. It does not claim this project's synthetic-voice detector, codec adaptations, or latency performance.
Tak, H., Todisco, M., Wang, X., Jung, J., Yamagishi, J., & Evans, N.2022
Automatic Speaker Verification Spoofing and Deepfake Detection Using Wav2vec 2.0 and Data Augmentation. Odyssey 2022, pp. 112-119.
Reports 0.82% LA EER and 2.85% DF EER for the paper's wav2vec 2.0 frontend, self-attentive aggregation, and augmentation setup. These literature results are not Noise or Voice measurements.