EU AI Act Article 50(2) applies from 2 August 2026. Providers of AI systems that generate synthetic audio must mark the output in a machine-readable format, detectable as artificially generated (Article 50 text, European Commission FAQ). The complete Fish Audio developer documentation corpus, 362,947 bytes, contains zero occurrences of watermark, C2PA, provenance, or content credential (llms-full.txt).
Hive comes alongside a voice stack as a sidecar. At the moment of each generation, each clone creation, and each takedown, it makes an independent, signed receipt that anyone can check offline. It never generates audio, never blocks a request, never decides whether a consent was valid, and never holds your keys. It is invisible to your users, invisible to your inference path, and fail-open. If Hive is unreachable, generation runs exactly as it does today.
This is the strongest fact on the page. It is not an opinion, a benchmark, or a competitive comparison. It is a date in force law, and it is days away.
Four days on the calendar from 29 July 2026. The counter above runs to 2 August 2026, 00:00 UTC, the date Article 50 transparency obligations begin to apply, so it reads in exact hours rather than calendar days.
Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated. EU AI Act, Article 50(2) · artificialintelligenceact.eu/article/50
Article 50 transparency obligations apply from 2 August 2026. Relief to 2 December 2026 is available only for systems already placed on the market before that date, so a system launched on 28 July 2026 is a live question rather than a settled one (European Commission FAQ).
Non-compliance with Article 50 carries administrative fines of up to EUR 15 million or 3 percent of total worldwide annual turnover, whichever is higher (European Commission FAQ).
The Commission's final Article 50 guidelines, adopted 20 July 2026, confirm that "persons" covers "digital replicas of real people... and personal characteristics or expressions such as a person's image, voice, behaviour and performances" (Bird & Bird).
Verified against the complete published developer documentation corpus, 362,947 bytes: docs.fish.audio/llms-full.txt. A term count is a fact about the published documentation. It is not a statement about what runs in production.
Media Origin Receipt™ and MoRSo™. A machine-readable, signed origin receipt emitted with every generation. Verifiable offline by anyone. No dependency on Fish Audio being available, reachable, or willing to confirm it. The mark travels with the file rather than living in a vendor database.
A receipt is not a Commission-approved marking method, and no regulator has reviewed it. It makes the origin of a file independently checkable. Whether a given marking approach satisfies Article 50(2) is a legal determination for Fish Audio's counsel, not one a receipt makes.
Handled factually. This is not an allegation of bad faith. It is an observation that two published Fish Audio surfaces do not currently line up, and that the gap is visible to anyone who reads both.
Cloned voices require consent verification. Audio watermarking is built into the cloning pipeline.
Consent attestation is part of the API call. Fish Audio developers page, verbatim · fish.audio/developers
The word consent appears zero times in the 363KB documentation corpus. So do watermark, C2PA, provenance, and content credential (llms-full.txt).
There is no consent field, parameter, or endpoint in the API reference or in the published OpenAPI schema (api.fish.audio/openapi.json). Authentication is a static bearer key (developers).
There is no published watermark specification, no verification endpoint, no robustness claim, and no third-party attestation to point a reviewer at (llms-full.txt).
A security reviewer at a regulated-industries buyer will read the marketing page, search the documentation for the feature it promised, fail to find it, and escalate. That escalation is not about the feature. It is about the difference between the two pages.
An unverifiable claim costs more in procurement than an absent feature, because it converts a product gap into a credibility question. A gap can be closed on a roadmap. A credibility question follows the vendor into every subsequent review.
Fish Audio does publish the enterprise posture a reviewer will also ask about, including that a SOC 2 Type II "audit is currently underway" (enterprise). The assurance table further down sets out what exists and what does not, in the company's own words.
R3Pv™ turns the claim into a checkable artifact. The consent attestation and the watermark assertion become one grouped, signed receipt that the buyer's own reviewer verifies offline, without taking anyone's word for it. The marketing sentence and the technical reality become the same object.
A receipt does not decide whether the original marketing sentence was accurate, does not certify a watermark as robust, and does not resolve a disclosure question. It fixes what was asserted, by which system, under which key, at which moment.
The published position places the rights question on the user. That is a defensible commercial choice. It is also a position with no artifact behind it, which is a different problem.
You are responsible for confirming you have the rights, consents, and disclosures required for any voice you clone... Fish Audio does not pre-clear individual use cases. Fish Audio voice-clone FAQ · fish.audio/voice-clone
The voice-clone FAQ says ten seconds is enough and that "most public-figure clips, podcast cuts, or phone-quality recordings work on the first try" (voice-clone). The documentation says "Never use celebrity or public figure voices without permission" (docs). Both are published. Only one can be the rule.
The model licence "incorporates an Acceptable Use Policy by reference" (LICENSE). No AUP is published. Both /acceptable-use/ and /aup/ return 404, and the phrase "Acceptable Use" appears zero times in the Terms (Terms). The binding document is the one nobody can read.
Equity, the UK performers' union, published a demand letter dated 22 May 2026 from Industrial Official Shannon Sailing on behalf of 22 members. It seeks removal of all unauthorised cloned voices plus lawful-basis and international-transfer disclosures under UK GDPR and the Data Protection Act 2018, and states the union was "yet to receive a reply" (Equity).
TechCrunch reported creator allegations of voices uploaded without consent and slow takedowns. Rissa Cao says takedowns are now automated in "less than three minutes" (TechCrunch).
A three-minute takedown is a real operational improvement. It is also a claim about a process, told after the fact, from the same systems whose conduct is in question. The union's letter asks for a lawful basis, not a response time.
OriginProof™ for human-origin attestation. SPR™ to prove the instruction came from an authorised party. Physiological Provenance Receipt™ to bind the reference audio to a live capture rather than a downloaded clip. A signed consent receipt cut at the moment the voice model is created, not reconstructed afterward from support tickets.
A receipt does not establish that consent was legally sufficient, does not resolve a rights dispute, and does not decide a UK GDPR lawful-basis question. It fixes the record of what was attested, by whom, and when.
Weights are downloadable from Hugging Face under cc-by-nc-sa-4.0 (s1-mini) and the fish-speech repository carries 31.7k stars (GitHub). Anyone can self-host and generate voices Fish Audio never sees.
The licence, the Fish Audio Research License, is free for research and non-commercial use only, requires a paid licence for commercial use, and requires "Built with Fish Audio" attribution (LICENSE). Model copyright sits with a second entity, 39 AI, INC. Reputation therefore attaches to output that Fish Audio cannot observe, under an attribution requirement that is currently an honour system.
A self-hosted deployment can honour the "Built with Fish Audio" requirement or quietly ignore it. Either way, the output carries the name in the eyes of anyone who hears it and no signal distinguishes the two cases.
S2S™ hardware-rooted GPU attestation and PBS™ provenance-bonded sandbox. A self-hosted deployment still emits a signed receipt proving which weights ran in which environment. MiR™ binds which model answered.
When an unattributed clip surfaces, the question stops being "was this yours" and becomes "is there a receipt". A receipted deployment can be separated from an unreceipted one without inspecting anyone's servers.
A receipt cannot compel an unlicensed self-hoster to emit one, cannot decide whether a licence was breached, and cannot attribute an unreceipted clip to any origin. It proves the environment and weights for deployments that opt in, which is what separates a compliant integrator from an unknown one.
In Lehrman v. Lovo, No. 1:24-cv-03770 (S.D.N.Y.), Judge Oetken's 10 July 2025 opinion let breach of contract, New York right-of-publicity, and New York consumer-protection claims proceed, while dismissing the Lanham Act and most copyright claims (Reuters, opinion PDF).
Lovo appears here only as litigation precedent. No comparison to any voice vendor is drawn or implied anywhere on this page.
Copyright does not protect against imitation of a voice, only direct copying of a fixed recording. Vendor exposure therefore runs through state right-of-publicity and contract, not federal IP (Loeb & Loeb, Skadden).
The court rejected a one-year limitations defence, reasoning that a model trained on the plaintiffs' voices "was arguably continuing to replicate those voices each time the model generated new clips" (opinion PDF). A trained model keeps the exposure current.
Lovo filed for bankruptcy and the copyright suit was automatically stayed in May 2026 (case tracker). The precedent survives the defendant.
Tennessee's ELVIS Act, in force 1 July 2024, reaches vendors directly. It creates liability for distributing "an algorithm, software, tool or other technology, service or device, the primary purpose or function of which is to produce a particular, identifiable individual's photograph, voice or likeness" without authorisation (Holland & Knight).
A tool-provider clause is what makes "the user agreed to our terms" an incomplete answer. The question becomes what the provider can show about the moment the voice model was made.
Forensic Rail™ for threshold-signed, deterministic replay under a consortium credential. Hive Ledger for append-only history. ViewKey™ so a rights-holder, a regulator, a union, and an enterprise customer each read the same event from their own vantage, without a shared database and without new access.
A receipt does not decide a right-of-publicity claim, does not satisfy a discovery obligation on its own, and does not make a filing go away. It replaces a self-authored log with an independently checkable record of what the system did.
Nine artifacts a regulated-industries security review asks for, with the current published status in Fish Audio's own words. One row is confirmed strong. Two are in progress or contract-gated. The rest are absences, and an absence is easier to close than a contradiction.
| Artifact | Fish Audio status | Source |
|---|---|---|
| SOC 2 Type II | "audit is currently underway... available to customers under NDA once it is complete" | enterprise |
| ISO 27001 | No mention anywhere | enterprise |
| HIPAA | "HIPAA-aligned configurations and can sign a BAA for qualifying healthcare workloads"; plan note "SOC2 / HIPAA upon requests" | enterprise |
| Zero data retention | "available on enterprise contracts", not the default; consumer policy retains | enterprise, Privacy |
| On-premises | Confirmed: "Deploy to your VPC, data center, sovereign cloud, or air-gapped environment" | developers |
| Trust / security page | Does not exist. /security/, /trust/, /compliance/ all 404 | probed |
| Published AUP | Does not exist, despite the licence incorporating one by reference | LICENSE |
| Watermarking | Marketing claim only. Zero occurrences in 363KB docs corpus | llms-full.txt |
| Consent attestation | Marketing claim only. Zero occurrences in docs corpus; no API field | llms-full.txt |
On a narrow screen, scroll the table sideways to read the source column. The on-premises row matters more than it looks. An air-gapped deployment is exactly the environment where a signed receipt is the only evidence that can leave, because the audio, the reference clip, and the logs cannot.
Nine rows. Every one names the primitive, the receipt it cuts, and the boundary it refuses to cross. The last line of each row is the point of the page. Hive proves conditions, never verdicts.
A machine-readable, signed origin receipt emitted with every generation, verifiable offline by anyone, with no dependency on Fish Audio being available or willing to confirm it. The mark is carried by the artifact rather than looked up in a vendor system.
It is not a Commission-approved marking method and no regulator has reviewed it. Whether a marking approach satisfies Article 50(2) is a legal determination for counsel, not one a receipt makes.
The consent attestation and the watermark assertion become one grouped, signed receipt that a buyer's own reviewer verifies offline. A marketing sentence becomes a checkable artifact, so the review ends at verification instead of escalation.
It does not judge whether the original claim was accurate, does not certify watermark robustness, and does not clear a disclosure question. It fixes what was asserted, by which system, under which key.
OriginProof™ attests human origin. SPR™ proves the instruction came from an authorised party. Physiological Provenance Receipt™ binds the reference audio to a live capture rather than a downloaded clip, which is the exact distinction a ten-second podcast cut erases.
It does not establish that consent was legally sufficient, does not resolve a rights dispute, and does not decide a UK GDPR lawful-basis question. It fixes the record of what was attested, by whom, and when.
S2S™ gives hardware-rooted GPU attestation. PBS™ runs the model in a provenance-bonded sandbox whose kernel, engine hash, and firmware version are on a signed heartbeat chain. MiR™ binds which model answered. Attribution stops being an honour system for deployments that opt in.
It cannot compel an unlicensed self-hoster to emit anything and cannot decide whether a licence was breached. It separates a receipted deployment from an unknown one.
Forensic Rail™ gives threshold-signed, deterministic replay under a consortium credential. Hive Ledger keeps append-only history. ViewKey™ lets a rights-holder, a regulator, a union, and an enterprise customer each read the same event from their own vantage, without a shared database and without new access.
It does not decide a right-of-publicity claim, does not discharge a discovery obligation, and does not adjudicate any filing. It replaces a self-authored log with an independently checkable record.
The thresholds that govern which voice uploads are refused, and which takedown conditions trigger removal, live on a public Merkle mutation ledger with a zero-knowledge envelope-bond. A union, a rights-holder, or an auditor can verify that the bounds were in force without seeing the exact numbers.
It does not decide whether a threshold was set correctly, adequately disclosed, or lawful. It proves which threshold was in force when a refusal or a removal ran.
A signed freeze receipt fires the instant an identifiable-likeness feature activates during a clone or generation request. The event is recorded at the moment it happens, not discovered later in a support thread.
It does not identify a person, does not decide whether a likeness was used without authorisation, and does not block the request. It records that a flagged feature activated.
Egress Bond™ meters caps on reference-audio and voice-identity egress per semantic class using Pedersen commitments, and a cap breach retroactively invalidates the downstream graph. Diurnal Bond™ requires k-of-n countersign before a high-impact action, such as reinstating a removed voice, runs during evening or weekend hours.
It does not decide what the correct cap should be, does not classify a transfer as lawful, and does not answer an international-transfer question. It proves the cap that applied and whether the second key signed.
Every receipt above folds into one append-only history, so the absence of an expected receipt is itself a signal. A certificate-transparency style log for receipt inclusion is designed, not yet serving. Nothing on this page depends on it, because verification is offline.
An append-only history cannot prove that an event nobody receipted ever happened. It can prove that a receipted event was never altered, and it can make a missing receipt visible.
Nothing above enters the inference path, refuses a generation, or decides a rights question. Each primitive proves that the thing that made a call was itself the thing it claimed to be. Hive proves conditions, never verdicts.
A receipt is only usable if it is free at the millisecond level. These are the measured figures, with the sample size and date attached, and with an honest label on maturity.
Sample size n=40. Measured 27 July 2026. 37 of 37 smoke tests pass. Figures are reported as medians and labelled as such.
Production integrations are pilot-ready rather than deployed. That is the accurate state today. The signing and verification path is measured and reproducible. The integrations into a live voice pipeline are not yet running in production anywhere, and this page does not claim they are.
The seven upstream pre-effect controls were filed as USPTO 64/119,279 on 26 July 2026. A filing is a filing. It is not a granted patent, and nothing on this page depends on it.
The certificate-transparency style inclusion log is designed, not yet serving. Verification does not depend on it. A receipt is checked offline against the signature and the content-addressed roots, with no call to Hive and no call to anyone else.
Signatures use a hybrid of Ed25519 and ML-DSA-65, the NIST FIPS 204 federal standard. Receipts carry one-way fingerprints and cryptographic commitments, never raw audio, never a reference clip, and never a voice embedding.
Numbers on this page are measurements of the Hive signing path, not projections about any Fish Audio system, and not a benchmark of any voice model.
In plain words: your organization holds the pen that signs. Hive never touches raw audio, never holds your keys, and never enters the generation path.
You keep the private key that stamps your receipts. Nobody else can sign in your name, and the key never leaves your own hardware. The signature is yours, not Hive's.
Hive gives you the way to make and check receipts. It does not generate audio, clone a voice, refuse a request, decide whether a consent was valid, or run your pipeline. Each event is recorded as a one-way fingerprint and a cryptographic commitment, never the underlying audio.
One surface. Not the platform, not a migration, not a rebuild. The obvious candidate is the generation endpoint, where a Media Origin Receipt™ can be emitted alongside the audio and checked offline by anyone. The second candidate is voice-model creation, where a consent receipt is cut at the moment the model is made. Hive builds it, runs it, and hands over receipts that verify without Hive in the loop. Nothing changes in the inference path. If the receipt rail is unreachable, generation proceeds exactly as it does today.
Naming Fish Audio does not imply any agreement. This page does not state or imply that Fish Audio is a customer, partner, pilot, or endorser, or that Fish Audio uses Hive. Customers referenced anywhere in Hive materials are HeyGen, LiveKit, Retell, and Telnyx integrations confirmed on the customer side (HeyGen, LiveKit, Retell, Telnyx), which are facts about Fish Audio's own integrations rather than about Hive.
One line to start: [email protected]
Stephen Rotzin, Founder, Hive Civilization Inc (Wyoming) · [email protected]
This page is illustrative only. It is prepared privately by Hive Civilization Inc. and is not affiliated with, sponsored by, or endorsed by Fish Audio, Hanabi AI Inc., 39 AI, INC., or any of their affiliates. It does not state or imply that Fish Audio is a customer, partner, pilot, or endorser, or that Fish Audio uses Hive.
Every statement about Fish Audio on this page is quoted or counted from a published Fish Audio surface or a named public source and linked to it. Nothing here is an accusation of wrongdoing. Where two Fish Audio surfaces differ, both are quoted and both are linked, and the difference is presented as a published fact rather than an inference about intent. Facts were verified on 29 July 2026 and should be re-verified before any external use.
A signed receipt proves that a step ran and what it returned. It does not decide legality, does not guarantee compliance with the EU AI Act or any other law, does not resolve a rights or consent dispute, and does not adjudicate any court matter. Lehrman v. Lovo is cited only as litigation precedent on the doctrinal question of voice imitation and vendor exposure. The Equity demand letter is a published demand, not a finding. Production integrations are pilot-ready rather than deployed, and the certificate-transparency style inclusion log is designed, not yet serving.
Each link opens that entry in the canon implementation explorer, where its schema, mint route, open verify route, auth requirement and implementation state are stated. The state shown here is read from the same registry file the explorer renders from, so the two cannot drift apart. Nothing here implies a customer, a deployment or an endorsement.