# Noise or Voice: Research and Product Context Noise or Voice is an acoustic-feature synthetic speech detector with a browser demonstration, a Python DSP engine, a Cloudflare REST pilot, and Python and TypeScript SDKs. The project publishes its small evaluation ledgers and exposes failure cases. The present evidence does not justify treating a verdict as definitive speaker authentication. ## What was measured - Golden regression set: 26 speech samples and 3 nonspeech edge cases. The 26 speech samples currently classify correctly. This set is selected for regression testing, not a challenge-wide benchmark. - Diverse demonstration set: 19 of 22 selected clips classify correctly at threshold 0.50. - Multi-dataset stress set: 168 selected speech clips, including 26 overlapping the golden set. The Python detector has 73.21% accuracy at threshold 0.50 and 33.98% internal score-based EER. The API JavaScript DSP code, run locally on the same files, has 73.21% accuracy and 35.41% EER; four threshold decisions differ. - In-the-Wild subset: 40 selected clips. Both implementations have 37.50% accuracy; Python EER is 60.00% and API JavaScript EER is 65.00%. This is a substantial domain-shift failure. - WaveFake subset: 64 ordered convenience samples, with 79.69% accuracy at threshold 0.50. The selected human LJSpeech clips have a high false-rejection rate. The subset selection is not representative or independent, and the source audio files are not included in the public Git checkout. The published score ledger includes audio hashes, per-sample scores, and manifest labels; consult the benchmark page for exact methods and limits. No full ASVspoof 2021 DF or LA partition result is claimed for this detector. Literature numbers use their authors' own protocols and are not an apples-to-apples ranking. ## Implementation and trust boundaries The detector computes acoustic features motivated by vocal production and waveform synthesis research. Measured features have overlapping distributions; none is an invariant human-versus-machine law. Channel changes, codec effects, speakers, recording style, and future synthesis methods can alter performance. The REST endpoint accepts uncompressed PCM WAV or explicitly labeled 16 kHz raw PCM16. It processes audio in memory. New audit rows record the audio SHA-256 digest, verdict, key prefix, and timing metadata; they omit client IP and user agent. The separate rate-limit table stores client IPs for bounded request counting, and historical audit rows may contain direct request identifiers. Retention terms still require operational policy. The browser demo analyzes audio locally and exports an unsigned observation summary. Optional Ed25519 receipts are vendor JSON records. A signature can establish integrity of the signed data relative to a separately trusted key, and an audio hash can bind the receipt to exact bytes. Neither establishes that the detector verdict is scientifically correct. The API exposes only a bare `attestation`, with no C2PA wrapper or compatibility option. WebSocket, gRPC, SIP, Twilio, Asterisk, FreeSWITCH, and Genesys examples in the documentation are proposed integration designs. The current SDK chunk helpers make repeated HTTP requests; they do not provide a persistent low-latency streaming service. ## Patent relationship US Patent 11,924,717 concerns communications analytics. Its claims do not establish this detector's efficacy. Deepfake detection algorithms are separate later research. ## Research references - ASVspoof 2021 challenge and evaluation protocols: https://www.asvspoof.org/index2021.html - WaveFake dataset and paper: https://github.com/joel-frank/wavefake - In-the-Wild generalization study: https://arxiv.org/abs/2203.16263 - F5-TTS conditional flow matching paper: https://aclanthology.org/2025.acl-long.313/ - C2PA technical specification: https://spec.c2pa.org/specifications/specifications/2.1/specs/C2PA_Specification.html ## Canonical pages - Home: https://noiseorvoice.com/ - Architecture: https://noiseorvoice.com/architecture/ - Benchmarks: https://noiseorvoice.com/benchmarks/ - Sample evaluation lab: https://noiseorvoice.com/benchmarks/proof/ - Biden-labeled sample pair (local labels, unverified source): https://noiseorvoice.com/forensics/biden/ - Obama-labeled sample pair (local labels, unverified source): https://noiseorvoice.com/forensics/obama/ - Trump-labeled sample pair (local labels, unverified source): https://noiseorvoice.com/forensics/trump/ - REST and SDK documentation: https://noiseorvoice.com/docs/ - Privacy and data handling: https://noiseorvoice.com/privacy/ - Terms of use: https://noiseorvoice.com/terms/ - Sitemap: https://noiseorvoice.com/sitemap.xml - Author: https://ciscocaceres.com - Research agenda: https://ciscocaceres.com/research/#noise-or-voice Public source availability is not established. Published ledgers and browser DSP are inspectable; the complete evaluation audio is not bundled for external reproduction. Compliance status is self-reported at https://noiseorvoice.com/data/trust-status.json and no independent assurance is established. The browser demo is local; server API audits, donated recordings, rate-limit IPs, and pilot emails have separate storage paths. No automatic contribution or audit deletion period is configured. See the privacy notice for receipt deletion and provider-copy limits. New contributions record versioned consent. Public SDK package distribution is not established. The deployed API and browser are currently unsigned; signed Python proof-ledger records are offline artifacts. Last reviewed: 2026-10-06 ## Separate ODSS research snapshot A previously frozen JavaScript DSP reference core; these historical results have not been rerun with the current detector source (ca6a48f4f408) completed an exploratory developer-run ODSS human/synthetic diagnostic on 64 recordings, including 32 test recordings from 4 speakers at unchanged threshold 0.50. It falsely flagged 10/16 human recordings and missed 4/16 synthetic recordings, with 0 processing errors and 0 abstentions. Balanced correct fraction was 56.25%; the descriptive speaker-cluster interval [40.625%, 78.125%] is fragile with only 4 clusters. Scores are uncalibrated. This developer-run diagnostic does not establish deployment readiness, authentication, streaming, radio or general-population performance. The original ASVspoof selection remains unchanged; its later completed diagnostic is reported separately below. Methods, licensed sources and protocol deviations: https://tuneharness.com/research/2026-10-06/#odss-two-class-diagnostic-false-flags-and-misses-retained Frozen aggregate: https://noiseorvoice.com/data/odss-exploratory-summary-20261006.json 6 October 2026 Pacific ASVspoof 5 research diagnostic: the historical frozen DSP snapshot (ca6a48f4f408) scored all 192 registered recordings at unchanged threshold 0.50. On the 96-record evaluation subset it falsely flagged 22/48 human recordings and missed 8/48 synthetic recordings; 66/96 were correct (68.75%), with 0 processing errors and 0 abstentions. This developer-selected, speaker-disjoint two-TTS-family diagnostic has no independent blinded custodian. Source/channel/training exposure and separate synthetic-content reuse scope remain unresolved. Scores are uncalibrated; these historical results do not validate the current deployed detector, authentication, streaming, radio use or general-population performance. Methods: https://tuneharness.com/research/2026-10-06/#asvspoof-5-two-class-diagnostic-completed-6-october-2026-pacific Aggregate: https://noiseorvoice.com/data/asv192-descriptive-summary-20261006-pacific.json Internal 18,000-trial ASVspoof 5 comparison (not blind). On 18,000 developer-selected, labeled trials from the ASVspoof 5 evaluation partition (2,000 human, 16,000 synthetic, attacks A17 to A32), every fixed rule was scored on every row: Worker 59.86% balanced and 81.18% raw, rejecting 1351/2000 human trials and missing 2036/16000 synthetic trials; learned feature model 55.52% balanced and 54.37% raw, rejecting 860/2000 human trials and missing 7353/16000 synthetic trials; author-released SSL model 59.07% balanced and 90.89% raw, rejecting 1637/2000 human trials and missing 2/16000 synthetic trials. The Worker abstained on 36 synthetic trials, which count as errors, and two retained human trials had prior developer exposure. Balanced accuracy is the primary figure because raw accuracy is inflated by the 8:1 synthetic share; all three balanced figures are near 60% or lower. Historical source scope: this comparison ran frozen DSP source fdfc11cd0cb2 (/reproduce/dsp-fdfc11cd0cb2456b364b4ae770549b0a85dcbd27c3bbbd91dbf0c732a8c16f8e.txt); this published ledger was not generated by the current source (afd2fa0e7600). Equal error rate (EER) uses only determinate rows and a boundary that varies with the data, so it is not an operating result. The denominators are 17964 of 18,000 rows for the Worker (its 36 abstentions have no score and are left out of the EER only) and 18000 for the learned model and for the SSL model. The SSL model has two EER definitions on the same 18000 rows. The spoof-margin EER (20.10%) ranks trials by spoof logit minus bona fide logit. The observed author bona fide logit EER (20.55%) ranks them by the raw author bona fide logit, negated so higher means more synthetic. The second is our own computation from the author-released model's raw output; it is not a figure the model's authors published. Adaptation and input differ: the learned model was fit on a study training subset, the Worker and learned model read the full WAV, and the SSL model read only the first 4.04 seconds. This is an internal, non-blind comparison with known labels. It is not independent, not a full or official ASVspoof 5 challenge result, and it establishes no superiority. The full audio is not bundled. Source: ASVspoof 5 database, DOI 10.5281/zenodo.14498691, Open Data Commons Attribution License. Summary: https://noiseorvoice.com/data/asv5-internal18k-summary-20261007.json Complete score ledger: https://noiseorvoice.com/data/asv5-internal18k-ledger-20261007.json