Embedded EU AI Act disclaimer
Every episode embeds a clear vocal notice stating it was produced with artificial intelligence. Not optional, not editable.
About — Editorial AI studio · 2026
Not a robotic reader of articles. Not a generic summariser. An on-demand audio studio that turns French-speaking curiosity into a directed listening experience — with editorial controls, signature voices and an EU compliance layer baked in.
Designed in France · Hosted in the EU · Native French-first
Our mission
Onde exists because audio production is the bottleneck of knowledge transmission. Articles get scrolled past. Reports get archived. Podcasts get scheduled but rarely match the exact topic, angle and duration you need today.
We do not summarise Wikipedia. We produce a directed episode: a real script, a real cast of voices, a real editorial intent. You stay the editor-in-chief — we run the studio.
Why French-first
NotebookLM, ElevenLabs GenFM and most American podcast generators were designed for English ears first, with French added later as a translated locale. Onde was built the other way around — and it shows.
ElevenLabs voices hand-picked for FR prosody and liaison accuracy, under commercial license.
Prompts written in French, French first-name bank tagged by tone and region, no English pivot.
Pension reform, EU AI Act, French jurisprudence — the model knows the references that matter for a French-speaking audience.
7 narrative genres with intro/outro jingles produced for a French audience — not stock libraries.
Editorial guardrails
Trust isn't a marketing claim. It's a stack of technical decisions you can verify in the audio file you download.
Every episode embeds a clear vocal notice stating it was produced with artificial intelligence. Not optional, not editable.
The MP3 file embeds a SHA-256 topic hash, episode ID, user ID and timestamp. End-to-end traceability without breaking the standard format.
An LLM filter blocks impersonation of politicians, journalists, executives, or reproduction of media brands (BFM, France Inter, etc.) before generation even starts.
Commercial ElevenLabs pool, never any non-consented cloning. Opt-in voice cloning will arrive with signed written consent, not before.
Automatic penalty on "Of course!", "Absolutely!", "Indeed,", "Great question", "Isn't it?". The adversarial scorer rejects scripts that feel robotic.
On sensitive topics (news, health, law, science, geopolitics), a Haiku classifier triggers a Sonnet agent equipped with web_search. Sources listed in the editorial pack.
Our stack
No technical posturing — just the choices that matter, and why we made them. The whole pipeline runs on European infrastructure.
Language model
Anthropic Claude — Haiku 4.5 (≤ 15 min) · Sonnet 4.6 (long form)
Haiku for speed and cost on short formats, Sonnet for narrative consistency on long-form episodes. The right model for the right cognitive budget.
Voice synthesis
ElevenLabs eleven_v3
The most natural voice model on the market in 2026. French voices hand-picked for prosodic consistency and accurate liaisons.
Audio assembly
ffmpeg with loudnorm −16 LUFS · TP −1.5 · LRA 11
Spotify, Apple Podcasts, Deezer standard. Distribution-ready out of the pipeline, no post-mortem touch-up required.
Object storage
Cloudflare R2
Zero egress fees, European infrastructure, low latency. Native choice for streamed audio.
Database
PostgreSQL via Supabase (EU region)
Data stored in the European Union. GDPR compliance by location, not by contractual workaround.
Web application
Next.js 16 · Vercel · EU hosting
Modern stack, European deployment, optimised runtime. No technical posturing: just the bricks that make the service fast and legible.
Editorial combinatorial
The same topic can become a pedagogical lecture or a documentary thriller, a neutral piece or a devil's advocate debate. That's a massive combinatorial space — and it's what turns the user into a director rather than a spectator.
Our approach vs the rest
Prompts are written in French, voices are native French, jingles are produced for the FR ear. No English pivot along the way.
7 formats × 13 tones × 13 angles = 1,183 possible editorial directions. The same topic can become a teaching lecture or a documentary thriller.
Mandatory vocal disclaimer, SHA-256 ID3 watermark, pre-generation moderation. Transparency is a product choice, not a concession.
Loudnorm −16 LUFS, signature jingles per genre, controlled transitions. The episode must hold up on headphones, in the car, on speakers.
Compliance commitments
Compliance isn't a checklist filed once a year. It's a daily operating constraint that shapes every product decision.
Public roadmap
We publish our priorities so you can hold us to them. Dates are targets, not contractual promises.
Q3 2026
English-speaking market opens with dedicated EN voices (no FR translation). Voice cloning with signed written consent for Pro creators.
Q4 2026
Documented REST API + webhooks to integrate Onde into an internal LMS (Cornerstone, Docebo, 360Learning). SAML / SCIM provisioning available on Enterprise.
2027
Reinforced grounding via curated corpus (Wikipedia + INSEE + Eurostat). White-label for large accounts and EdTech platforms.
Dig deeper
The five pipeline steps, from prompt to MP3, with the studio control-room animation.
Onde vs NotebookLM, ElevenLabs GenFM, Wondercraft — fifteen criteria detailed.
Onboarding, compliance, market intel, knowledge capture, upskilling — Team plan at €199/month.
Free plan (15 credits / month) · Starter at €14 · Pro at €39 · Team B2B from €199.
The studio is open. Free-form topic, editorial controls, an episode in under 3 minutes. 15 free credits to test, then €14 a month if you stay.
Contact us: hello@ondeaudio.com