Anyone can generate a translated voice track. Toluva can show you exactly how it was produced, who approved it, and what changed when it did not fit.
Voice authorization is an active generation control, not metadata. Language, purpose, validity window, and the consent-evidence hash are all checked before a provider is ever called.
Every translated segment is measured against its source slot. Drift outside the band triggers a bounded rewrite and regeneration loop — red to green, with each attempt kept.
Backblaze B2 holds source masters, transcripts, every speech attempt, captions, renders, and canonical Genblaze manifests. Stored hashes are re-checked before bytes are called verified.
One 3.800s source slot · protected term Toluva
Willkommen bei Toluva. Mit unserer Plattform wird eine einzige…
Attempt 1 · blockedWillkommen bei Toluva. Eine Botschaft, viele Sprachen.
Attempt 2 · approved rewriteThe German for one 3.8-second slot generated 8.127 seconds of speech. Toluva measured it, blocked the next billable call, and waited for a hash-bound human-approved rewrite. The shorter revision generated 3.576 seconds — a −5.9% fit. Both attempts are still on the record.
Inspect the correction proofEvery generation request is evaluated against the stored authorization — permitted languages, approved purposes, validity window, revocation, and the evidence hash. A request outside that scope is refused before a single credit is spent, and the refusal is recorded too.
Test the boundary369f3eea954c2bba91bd7a65cade78a86a9f9e1050cf915702e9a2da2e3917fe
Transcription, three translations, three speech runs, the source-timed audio fan-in, and the three-input video composition. Each is its own run with its own canonical manifest, and the app re-checks stored media hashes before calling any of it verified.
Sixty job-scoped objects for a single twelve-second run. Source masters, authorization evidence, quality decisions, every speech attempt, disclosure records, and canonical manifests all land in B2 before anything is called finished.
Failed attempts are never overwritten or hidden. Each retry is a separate run carrying the previous attempt's parent run ID, so the whole correction history stays auditable.
Public routes refuse writes before upload, approval, or provider spend. Credentials stay server-side, and media proxies accept fixed kinds or opaque job handles — never arbitrary B2 keys.
A manifest proves recorded lineage and canonical integrity. It does not prove every supplied fact or guarantee regulatory compliance, and Toluva does not claim otherwise.
The submitted deployment proves one complete English-to-German production lane. It does not claim universal language support, perfect lip sync, or automatic legal compliance.
A controlled English source processed into three timed segments, then recomposed as H.264 video with AAC audio, embedded mov_text captions, and a WebVTT sidecar.
Source lineage, authorization evidence, transcripts, quality decisions, every speech attempt, captions, renders, and canonical manifests — all job-scoped.
Transcription, three translations, three speech runs, the audio fan-in, and the final composition. Every one independently hash-checked before display.
One German attempt overran its 3.800s slot by 113.868%. A protected-term-safe, human-approved rewrite landed at −5.898%. Both attempts remain inspectable.
New jobs use a bounded German lane with rights confirmation, synthetic-voice disclosure, daily admission slots, and a hard per-job provider budget.